[llvm] AMDGPU: Cleanup FP atomicrmw tests and cover fmin/fmax (PR #95131)
Matt Arsenault via llvm-commits
llvm-commits at lists.llvm.org
Thu Jun 13 05:41:44 PDT 2024
https://github.com/arsenm updated https://github.com/llvm/llvm-project/pull/95131
>From ef7115028a28b7817c33d6079b17e0d56fde063f Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Tue, 11 Jun 2024 14:50:04 +0200
Subject: [PATCH 1/4] AMDGPU: Add codegen tests for local atomicrmw fmin/fmax
Also make the test organization more consistent.
---
.../CodeGen/AMDGPU/local-atomicrmw-fadd.ll | 9102 +++++++++++++++++
.../CodeGen/AMDGPU/local-atomicrmw-fmax.ll | 8456 +++++++++++++++
.../CodeGen/AMDGPU/local-atomicrmw-fmin.ll | 8456 +++++++++++++++
.../CodeGen/AMDGPU/local-atomicrmw-fsub.ll | 8226 +++++++++++++++
llvm/test/CodeGen/AMDGPU/local-atomics-fp.ll | 3931 -------
5 files changed, 34240 insertions(+), 3931 deletions(-)
create mode 100644 llvm/test/CodeGen/AMDGPU/local-atomicrmw-fadd.ll
create mode 100644 llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmax.ll
create mode 100644 llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmin.ll
create mode 100644 llvm/test/CodeGen/AMDGPU/local-atomicrmw-fsub.ll
delete mode 100644 llvm/test/CodeGen/AMDGPU/local-atomics-fp.ll
diff --git a/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fadd.ll b/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fadd.ll
new file mode 100644
index 0000000000000..4eaa4274a1eac
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fadd.ll
@@ -0,0 +1,9102 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 < %s | FileCheck -check-prefix=GFX12 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx940 < %s | FileCheck -check-prefix=GFX940 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx90a < %s | FileCheck -check-prefix=GFX90A %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx908 < %s | FileCheck -check-prefix=GFX908 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tonga < %s | FileCheck -check-prefix=GFX8 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=hawaii < %s | FileCheck -check-prefix=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tahiti < %s | FileCheck -check-prefix=GFX6 %s
+
+; --------------------------------------------------------------------
+; float
+; --------------------------------------------------------------------
+
+define float @local_atomic_fadd_ret_f32(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fadd_ret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_add_rtn_f32 v0, v0, v1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_ret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX940-NEXT: ds_add_rtn_f32 v0, v0, v1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_ret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_add_rtn_f32 v0, v0, v1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_ret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_add_rtn_f32 v0, v0, v1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_ret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX90A-NEXT: ds_add_rtn_f32 v0, v0, v1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_ret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX908-NEXT: ds_add_rtn_f32 v0, v0, v1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_ret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_add_rtn_f32 v0, v0, v1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_ret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v1
+; GFX7-NEXT: v_add_f32_e32 v1, 4.0, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB0_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_ret_f32:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v1
+; GFX6-NEXT: v_add_f32_e32 v1, 4.0, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB0_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v0, v1
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(3) %ptr, float 4.0 seq_cst
+ ret float %result
+}
+
+define float @local_atomic_fadd_ret_f32__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fadd_ret_f32__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_add_rtn_f32 v0, v0, v1 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_ret_f32__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX940-NEXT: ds_add_rtn_f32 v0, v0, v1 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_ret_f32__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_add_rtn_f32 v0, v0, v1 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_ret_f32__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_add_rtn_f32 v0, v0, v1 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_ret_f32__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX90A-NEXT: ds_add_rtn_f32 v0, v0, v1 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_ret_f32__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX908-NEXT: ds_add_rtn_f32 v0, v0, v1 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_ret_f32__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_add_rtn_f32 v0, v0, v1 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_ret_f32__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v1
+; GFX7-NEXT: v_add_f32_e32 v1, 4.0, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB1_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_ret_f32__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v1, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v0, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v0
+; GFX6-NEXT: v_add_f32_e32 v0, 4.0, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v0, v1, v2, v0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v0, v2
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB1_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fadd ptr addrspace(3) %gep, float 4.0 seq_cst
+ ret float %result
+}
+
+define void @local_atomic_fadd_noret_f32(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fadd_noret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_add_f32 v0, v1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_noret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX940-NEXT: ds_add_f32 v0, v1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_noret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_add_f32 v0, v1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_noret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_add_f32 v0, v1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_noret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX90A-NEXT: ds_add_f32 v0, v1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_noret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX908-NEXT: ds_add_f32 v0, v1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_noret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_add_f32 v0, v1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_noret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v2, 4.0, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_noret_f32:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v2, 4.0, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB2_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(3) %ptr, float 4.0 seq_cst
+ ret void
+}
+
+define void @local_atomic_fadd_noret_f32__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fadd_noret_f32__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_add_f32 v0, v1 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_noret_f32__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX940-NEXT: ds_add_f32 v0, v1 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_noret_f32__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_add_f32 v0, v1 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_noret_f32__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_add_f32 v0, v1 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_noret_f32__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX90A-NEXT: ds_add_f32 v0, v1 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_noret_f32__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX908-NEXT: ds_add_f32 v0, v1 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_noret_f32__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_add_f32 v0, v1 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_noret_f32__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v2, 4.0, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB3_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_noret_f32__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v2, 4.0, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB3_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(3) %ptr, i32 16383
+ %unused = atomicrmw fadd ptr addrspace(3) %gep, float 4.0 seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; double
+; --------------------------------------------------------------------
+
+define double @local_atomic_fadd_ret_f64(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fadd_ret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v0
+; GFX12-NEXT: ds_load_b64 v[0:1], v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_add_f64_e32 v[0:1], 4.0, v[3:4]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4]
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB4_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_ret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], 4.0
+; GFX940-NEXT: ds_add_rtn_f64 v[0:1], v0, v[2:3]
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_ret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v0
+; GFX11-NEXT: ds_load_b64 v[0:1], v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_f64 v[0:1], v[3:4], 4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4]
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB4_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_ret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v0
+; GFX10-NEXT: ds_read_b64 v[0:1], v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v1
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_add_f64 v[0:1], v[3:4], 4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB4_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_ret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, 0
+; GFX90A-NEXT: v_mov_b32_e32 v3, 0x40100000
+; GFX90A-NEXT: ds_add_rtn_f64 v[0:1], v0, v[2:3]
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_ret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v0
+; GFX908-NEXT: ds_read_b64 v[0:1], v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v1
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_add_f64 v[0:1], v[3:4], 4.0
+; GFX908-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB4_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_ret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: v_mov_b32_e32 v2, v0
+; GFX8-NEXT: ds_read_b64 v[0:1], v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v1
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_add_f64 v[0:1], v[3:4], 4.0
+; GFX8-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB4_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_ret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: v_mov_b32_e32 v2, v0
+; GFX7-NEXT: ds_read_b64 v[0:1], v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_add_f64 v[0:1], v[3:4], 4.0
+; GFX7-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB4_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_ret_f64:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: v_mov_b32_e32 v2, v0
+; GFX6-NEXT: ds_read_b64 v[0:1], v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: v_add_f64 v[0:1], v[3:4], 4.0
+; GFX6-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB4_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(3) %ptr, double 4.0 seq_cst
+ ret double %result
+}
+
+define double @local_atomic_fadd_ret_f64__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fadd_ret_f64__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v0
+; GFX12-NEXT: ds_load_b64 v[0:1], v0 offset:65528
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_add_f64_e32 v[0:1], 4.0, v[3:4]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4] offset:65528
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB5_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_ret_f64__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], 4.0
+; GFX940-NEXT: ds_add_rtn_f64 v[0:1], v0, v[2:3] offset:65528
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_ret_f64__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v0
+; GFX11-NEXT: ds_load_b64 v[0:1], v0 offset:65528
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_f64 v[0:1], v[3:4], 4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4] offset:65528
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB5_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_ret_f64__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v0
+; GFX10-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v1
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_add_f64 v[0:1], v[3:4], 4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] offset:65528
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB5_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_ret_f64__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, 0
+; GFX90A-NEXT: v_mov_b32_e32 v3, 0x40100000
+; GFX90A-NEXT: ds_add_rtn_f64 v[0:1], v0, v[2:3] offset:65528
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_ret_f64__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v0
+; GFX908-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v1
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_add_f64 v[0:1], v[3:4], 4.0
+; GFX908-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] offset:65528
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_ret_f64__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: v_mov_b32_e32 v2, v0
+; GFX8-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v1
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_add_f64 v[0:1], v[3:4], 4.0
+; GFX8-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] offset:65528
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_ret_f64__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: v_mov_b32_e32 v2, v0
+; GFX7-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_add_f64 v[0:1], v[3:4], 4.0
+; GFX7-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] offset:65528
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_ret_f64__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v2, vcc, 0xfff8, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b64 v[0:1], v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: v_add_f64 v[0:1], v[3:4], 4.0
+; GFX6-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB5_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(3) %ptr, i32 8191
+ %result = atomicrmw fadd ptr addrspace(3) %gep, double 4.0 seq_cst
+ ret double %result
+}
+
+define void @local_atomic_fadd_noret_f64(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fadd_noret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b64 v[1:2], v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_add_f64_e32 v[3:4], 4.0, v[1:2]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2]
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX12-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB6_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_noret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], 4.0
+; GFX940-NEXT: ds_add_f64 v0, v[2:3]
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_noret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b64 v[1:2], v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_add_f64 v[3:4], v[1:2], 4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2]
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX11-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB6_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_noret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b64 v[1:2], v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_add_f64 v[3:4], v[1:2], 4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX10-NEXT: v_mov_b32_e32 v1, v3
+; GFX10-NEXT: v_mov_b32_e32 v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB6_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_noret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, 0
+; GFX90A-NEXT: v_mov_b32_e32 v3, 0x40100000
+; GFX90A-NEXT: ds_add_f64 v0, v[2:3]
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_noret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b64 v[1:2], v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_add_f64 v[3:4], v[1:2], 4.0
+; GFX908-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX908-NEXT: v_mov_b32_e32 v1, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v2, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB6_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_noret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b64 v[1:2], v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_add_f64 v[3:4], v[1:2], 4.0
+; GFX8-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX8-NEXT: v_mov_b32_e32 v1, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v2, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB6_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_noret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b64 v[1:2], v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_add_f64 v[3:4], v[1:2], 4.0
+; GFX7-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v2, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB6_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_noret_f64:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b64 v[1:2], v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_add_f64 v[3:4], v[1:2], 4.0
+; GFX6-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX6-NEXT: v_mov_b32_e32 v1, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v2, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB6_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(3) %ptr, double 4.0 seq_cst
+ ret void
+}
+
+define void @local_atomic_fadd_noret_f64__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fadd_noret_f64__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b64 v[1:2], v0 offset:65528
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_add_f64_e32 v[3:4], 4.0, v[1:2]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2] offset:65528
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX12-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB7_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_noret_f64__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], 4.0
+; GFX940-NEXT: ds_add_f64 v0, v[2:3] offset:65528
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_noret_f64__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b64 v[1:2], v0 offset:65528
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_add_f64 v[3:4], v[1:2], 4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2] offset:65528
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX11-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB7_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_noret_f64__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b64 v[1:2], v0 offset:65528
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_add_f64 v[3:4], v[1:2], 4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX10-NEXT: v_mov_b32_e32 v1, v3
+; GFX10-NEXT: v_mov_b32_e32 v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB7_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_noret_f64__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, 0
+; GFX90A-NEXT: v_mov_b32_e32 v3, 0x40100000
+; GFX90A-NEXT: ds_add_f64 v0, v[2:3] offset:65528
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_noret_f64__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b64 v[1:2], v0 offset:65528
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_add_f64 v[3:4], v[1:2], 4.0
+; GFX908-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX908-NEXT: v_mov_b32_e32 v1, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v2, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB7_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_noret_f64__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b64 v[1:2], v0 offset:65528
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_add_f64 v[3:4], v[1:2], 4.0
+; GFX8-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX8-NEXT: v_mov_b32_e32 v1, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v2, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB7_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_noret_f64__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b64 v[1:2], v0 offset:65528
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_add_f64 v[3:4], v[1:2], 4.0
+; GFX7-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v2, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB7_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_noret_f64__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v2, vcc, 0xfff8, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b64 v[0:1], v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_add_f64 v[3:4], v[0:1], 4.0
+; GFX6-NEXT: ds_cmpst_rtn_b64 v[3:4], v2, v[0:1], v[3:4]
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[0:1]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB7_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(3) %ptr, i32 8191
+ %unused = atomicrmw fadd ptr addrspace(3) %gep, double 4.0 seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; half
+; --------------------------------------------------------------------
+
+define half @local_atomic_fadd_ret_f16(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fadd_ret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: ds_load_b32 v2, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX12-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_not_b32_e32 v3, v3
+; GFX12-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v0, v4
+; GFX12-NEXT: v_add_f16_e32 v2, 4.0, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v0, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v1, v2, v4
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB8_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v0, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_ret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v2, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_and_b32_e32 v0, 24, v3
+; GFX940-NEXT: v_lshlrev_b32_e64 v3, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v3, v3
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v2
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, v0, v4
+; GFX940-NEXT: v_add_f16_e32 v2, 4.0, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, v0, v2
+; GFX940-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v1, v4, v2
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB8_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v0, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_ret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: ds_load_b32 v2, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX11-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_not_b32_e32 v3, v3
+; GFX11-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v0, v4
+; GFX11-NEXT: v_add_f16_e32 v2, 4.0, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v0, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v1, v2, v4
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB8_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v0, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_ret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: ds_read_b32 v2, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX10-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_not_b32_e32 v3, v3
+; GFX10-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, v0, v4
+; GFX10-NEXT: v_add_f16_e32 v2, 4.0, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v1, v4, v2
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB8_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_ret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v2, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_and_b32_e32 v0, 24, v3
+; GFX90A-NEXT: v_lshlrev_b32_e64 v3, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v3, v3
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v2
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, v0, v4
+; GFX90A-NEXT: v_add_f16_e32 v2, 4.0, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, v0, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v1, v4, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB8_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_ret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v2, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_and_b32_e32 v0, 24, v3
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v3, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, v0, v4
+; GFX908-NEXT: v_add_f16_e32 v2, 4.0, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, v0, v2
+; GFX908-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v1, v4, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v0, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_ret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX8-NEXT: v_add_f16_e32 v3, 4.0, v3
+; GFX8-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_ret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX7-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB8_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_ret_f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX6-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX6-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB8_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(3) %ptr, half 4.0 seq_cst
+ ret half %result
+}
+
+define half @local_atomic_fadd_ret_f16__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fadd_ret_f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX12-NEXT: ds_load_b32 v3, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX12-NEXT: v_add_f16_e32 v3, 4.0, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB9_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_ret_f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v2, v1
+; GFX940-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v3, v0, s0
+; GFX940-NEXT: v_not_b32_e32 v3, v3
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v2
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, v0, v4
+; GFX940-NEXT: v_add_f16_e32 v2, 4.0, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, v0, v2
+; GFX940-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v1, v4, v2
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB9_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v0, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_ret_f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX11-NEXT: ds_load_b32 v3, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX11-NEXT: v_add_f16_e32 v3, 4.0, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB9_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_ret_f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX10-NEXT: ds_read_b32 v2, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v1, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v3, v3
+; GFX10-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, v1, v4
+; GFX10-NEXT: v_add_f16_e32 v2, 4.0, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB9_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v1, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_ret_f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v2, v1
+; GFX90A-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v3, v0, s4
+; GFX90A-NEXT: v_not_b32_e32 v3, v3
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v2
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, v0, v4
+; GFX90A-NEXT: v_add_f16_e32 v2, 4.0, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, v0, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v1, v4, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB9_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_ret_f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v2, v1
+; GFX908-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v0, s4
+; GFX908-NEXT: v_not_b32_e32 v3, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, v0, v4
+; GFX908-NEXT: v_add_f16_e32 v2, 4.0, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, v0, v2
+; GFX908-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v1, v4, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB9_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v0, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_ret_f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffe, v0
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v0, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX8-NEXT: v_add_f16_e32 v3, 4.0, v3
+; GFX8-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB9_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_ret_f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX7-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB9_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_ret_f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX6-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB9_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(3) %ptr, i32 32767
+ %result = atomicrmw fadd ptr addrspace(3) %gep, half 4.0 seq_cst
+ ret half %result
+}
+
+define void @local_atomic_fadd_noret_f16(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fadd_noret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: ds_load_b32 v2, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX12-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_not_b32_e32 v3, v3
+; GFX12-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX12-NEXT: v_add_f16_e32 v4, 4.0, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX12-NEXT: v_mov_b32_e32 v2, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB10_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_noret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v2, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX940-NEXT: v_add_f16_e32 v4, 4.0, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX940-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB10_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_noret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: ds_load_b32 v2, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX11-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_not_b32_e32 v3, v3
+; GFX11-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX11-NEXT: v_add_f16_e32 v4, 4.0, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX11-NEXT: v_mov_b32_e32 v2, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB10_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_noret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: ds_read_b32 v2, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX10-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_not_b32_e32 v3, v3
+; GFX10-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX10-NEXT: v_add_f16_e32 v4, 4.0, v4
+; GFX10-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v1, v2, v4
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX10-NEXT: v_mov_b32_e32 v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB10_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_noret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX90A-NEXT: v_add_f16_e32 v4, 4.0, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB10_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_noret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX908-NEXT: v_add_f16_e32 v4, 4.0, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB10_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_noret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX8-NEXT: v_add_f16_e32 v4, 4.0, v4
+; GFX8-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB10_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_noret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX7-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB10_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_noret_f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX6-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX6-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX6-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB10_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(3) %ptr, half 4.0 seq_cst
+ ret void
+}
+
+define void @local_atomic_fadd_noret_f16__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fadd_noret_f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX12-NEXT: ds_load_b32 v3, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f16_e32 v4, 4.0, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX12-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB11_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_noret_f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX940-NEXT: ds_read_b32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v1, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX940-NEXT: v_add_f16_e32 v4, 4.0, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX940-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB11_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_noret_f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX11-NEXT: ds_load_b32 v3, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f16_e32 v4, 4.0, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX11-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB11_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_noret_f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX10-NEXT: ds_read_b32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v2, v2
+; GFX10-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX10-NEXT: v_add_f16_e32 v4, 4.0, v4
+; GFX10-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB11_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_noret_f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX90A-NEXT: ds_read_b32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX90A-NEXT: v_add_f16_e32 v4, 4.0, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_noret_f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX908-NEXT: ds_read_b32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX908-NEXT: v_add_f16_e32 v4, 4.0, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_noret_f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0xfffe, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX8-NEXT: v_add_f16_e32 v4, 4.0, v4
+; GFX8-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_noret_f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX7-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB11_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_noret_f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX6-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX6-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB11_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(3) %ptr, i32 32767
+ %unused = atomicrmw fadd ptr addrspace(3) %gep, half 4.0 seq_cst
+ ret void
+}
+
+define half @local_atomic_fadd_ret_f16__offset__align4(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fadd_ret_f16__offset__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f16_e32 v1, 4.0, v2
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB12_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_ret_f16__offset__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v1
+; GFX940-NEXT: v_add_f16_e32 v1, 4.0, v2
+; GFX940-NEXT: v_and_or_b32 v1, v2, s2, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB12_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_ret_f16__offset__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f16_e32 v1, 4.0, v2
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB12_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_ret_f16__offset__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v1
+; GFX10-NEXT: v_add_f16_e32 v1, 4.0, v2
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX10-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB12_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v1
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_ret_f16__offset__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v1
+; GFX90A-NEXT: v_add_f16_e32 v1, 4.0, v2
+; GFX90A-NEXT: v_and_or_b32 v1, v2, s6, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB12_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_ret_f16__offset__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v1
+; GFX908-NEXT: v_add_f16_e32 v1, 4.0, v2
+; GFX908-NEXT: v_and_or_b32 v1, v2, s6, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB12_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_ret_f16__offset__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v1
+; GFX8-NEXT: v_add_f16_e32 v1, 4.0, v2
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX8-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB12_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_ret_f16__offset__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX7-NEXT: v_add_f32_e32 v1, 4.0, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB12_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_ret_f16__offset__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffe, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX6-NEXT: v_add_f32_e32 v1, 4.0, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB12_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(3) %ptr, i32 32767
+ %result = atomicrmw fadd ptr addrspace(3) %gep, half 4.0 seq_cst, align 4
+ ret half %result
+}
+
+define void @local_atomic_fadd_noret_f16__offset__align4(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fadd_noret_f16__offset__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_add_f16_e32 v2, 4.0, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB13_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_noret_f16__offset__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_add_f16_e32 v2, 4.0, v1
+; GFX940-NEXT: v_and_or_b32 v2, v1, s2, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB13_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_noret_f16__offset__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_add_f16_e32 v2, 4.0, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB13_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_noret_f16__offset__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_add_f16_e32 v2, 4.0, v1
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB13_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_noret_f16__offset__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_add_f16_e32 v2, 4.0, v1
+; GFX90A-NEXT: v_and_or_b32 v2, v1, s6, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB13_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_noret_f16__offset__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_add_f16_e32 v2, 4.0, v1
+; GFX908-NEXT: v_and_or_b32 v2, v1, s6, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB13_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_noret_f16__offset__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_add_f16_e32 v2, 4.0, v1
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX8-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB13_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_noret_f16__offset__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX7-NEXT: v_add_f32_e32 v2, 4.0, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB13_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_noret_f16__offset__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffe, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v1
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX6-NEXT: v_add_f32_e32 v2, 4.0, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB13_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(3) %ptr, i32 32767
+ %unused = atomicrmw fadd ptr addrspace(3) %gep, half 4.0 seq_cst, align 4
+ ret void
+}
+
+; --------------------------------------------------------------------
+; bfloat
+; --------------------------------------------------------------------
+
+define bfloat @local_atomic_fadd_ret_bf16(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fadd_ret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: ds_load_b32 v3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff
+; GFX12-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX12-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB14_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_ret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v2, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v3
+; GFX940-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX940-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX940-NEXT: v_add3_u32 v5, v5, v3, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB14_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_ret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: ds_load_b32 v3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff
+; GFX11-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX11-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB14_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_ret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: ds_read_b32 v3, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff
+; GFX10-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_not_b32_e32 v2, v2
+; GFX10-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX10-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX10-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB14_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_ret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v3
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX90A-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX90A-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_ret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX908-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_ret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX8-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v3
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v3, v6, v7, vcc
+; GFX8-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX8-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_ret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_ret_bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX6-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB14_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(3) %ptr, bfloat 4.0 seq_cst
+ ret bfloat %result
+}
+
+define bfloat @local_atomic_fadd_ret_bf16__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fadd_ret_bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX12-NEXT: ds_load_b32 v3, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX12-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB15_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_ret_bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v3, v1
+; GFX940-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v0, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v3
+; GFX940-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX940-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX940-NEXT: v_add3_u32 v5, v5, v3, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB15_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_ret_bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX11-NEXT: ds_load_b32 v3, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX11-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB15_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_ret_bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX10-NEXT: ds_read_b32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v2, v2
+; GFX10-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_lshrrev_b32_sdwa v3, v1, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX10-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX10-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB15_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_ret_bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v3, v1
+; GFX90A-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v0, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v3
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX90A-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX90A-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB15_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_ret_bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v3, v1
+; GFX908-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v0, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX908-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB15_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_ret_bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffe, v0
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v0, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX8-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v3
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v3, v6, v7, vcc
+; GFX8-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX8-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB15_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_ret_bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB15_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_ret_bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_add_f32_e32 v3, 4.0, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB15_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(3) %ptr, i32 32767
+ %result = atomicrmw fadd ptr addrspace(3) %gep, bfloat 4.0 seq_cst
+ ret bfloat %result
+}
+
+define void @local_atomic_fadd_noret_bf16(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fadd_noret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: ds_load_b32 v2, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX12-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_not_b32_e32 v3, v3
+; GFX12-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX12-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX12-NEXT: v_mov_b32_e32 v2, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB16_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_noret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v2, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX940-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX940-NEXT: v_add3_u32 v5, v5, v4, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB16_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_noret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: ds_load_b32 v2, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX11-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_not_b32_e32 v3, v3
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX11-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX11-NEXT: v_mov_b32_e32 v2, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB16_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_noret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: ds_read_b32 v2, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX10-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_not_b32_e32 v3, v3
+; GFX10-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v4, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX10-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX10-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v1, v2, v4
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX10-NEXT: v_mov_b32_e32 v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB16_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_noret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX90A-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX90A-NEXT: v_add3_u32 v5, v5, v4, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB16_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_noret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX908-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX908-NEXT: v_add3_u32 v5, v5, v4, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX908-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB16_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_noret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v4, v6, v7, vcc
+; GFX8-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX8-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB16_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_noret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB16_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_noret_bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX6-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX6-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB16_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(3) %ptr, bfloat 4.0 seq_cst
+ ret void
+}
+
+define void @local_atomic_fadd_noret_bf16__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fadd_noret_bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX12-NEXT: ds_load_b32 v3, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX12-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX12-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB17_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_noret_bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX940-NEXT: ds_read_b32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v1, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX940-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX940-NEXT: v_add3_u32 v5, v5, v4, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB17_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_noret_bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX11-NEXT: ds_load_b32 v3, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX11-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB17_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_noret_bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX10-NEXT: ds_read_b32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v2, v2
+; GFX10-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX10-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX10-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB17_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_noret_bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX90A-NEXT: ds_read_b32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX90A-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX90A-NEXT: v_add3_u32 v5, v5, v4, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB17_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_noret_bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX908-NEXT: ds_read_b32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX908-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX908-NEXT: v_add3_u32 v5, v5, v4, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX908-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_noret_bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0xfffe, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v4, v6, v7, vcc
+; GFX8-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX8-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_noret_bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_noret_bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_add_f32_e32 v4, 4.0, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX6-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB17_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(3) %ptr, i32 32767
+ %unused = atomicrmw fadd ptr addrspace(3) %gep, bfloat 4.0 seq_cst
+ ret void
+}
+
+define bfloat @local_atomic_fadd_ret_bf16__offset__align4(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fadd_ret_bf16__offset__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX12-NEXT: v_add_f32_e32 v1, 4.0, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX12-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB18_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_ret_bf16__offset__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX940-NEXT: v_add_f32_e32 v1, 4.0, v1
+; GFX940-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX940-NEXT: v_add3_u32 v3, v3, v1, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX940-NEXT: v_and_or_b32 v1, v2, s3, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB18_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_ret_bf16__offset__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX11-NEXT: v_add_f32_e32 v1, 4.0, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB18_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_ret_bf16__offset__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v1
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX10-NEXT: v_add_f32_e32 v1, 4.0, v1
+; GFX10-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX10-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX10-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB18_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v1
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_ret_bf16__offset__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX90A-NEXT: v_add_f32_e32 v1, 4.0, v1
+; GFX90A-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX90A-NEXT: v_add3_u32 v3, v3, v1, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX90A-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX90A-NEXT: v_and_or_b32 v1, v2, s7, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB18_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_ret_bf16__offset__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX908-NEXT: v_add_f32_e32 v1, 4.0, v1
+; GFX908-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX908-NEXT: v_add3_u32 v3, v3, v1, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX908-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX908-NEXT: v_and_or_b32 v1, v2, s7, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB18_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_ret_bf16__offset__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_add_f32_e32 v1, 4.0, v1
+; GFX8-NEXT: v_bfe_u32 v4, v1, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v1
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
+; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v1
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v1, v4, v5, vcc
+; GFX8-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB18_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_ret_bf16__offset__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX7-NEXT: v_add_f32_e32 v1, 4.0, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB18_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_ret_bf16__offset__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffe, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX6-NEXT: v_add_f32_e32 v1, 4.0, v1
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB18_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(3) %ptr, i32 32767
+ %result = atomicrmw fadd ptr addrspace(3) %gep, bfloat 4.0 seq_cst, align 4
+ ret bfloat %result
+}
+
+define void @local_atomic_fadd_noret_bf16__offset__align4(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fadd_noret_bf16__offset__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v2, 4.0, v2
+; GFX12-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB19_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_noret_bf16__offset__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX940-NEXT: v_add_f32_e32 v2, 4.0, v2
+; GFX940-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v3, v3, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: v_and_or_b32 v2, v1, s3, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB19_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_noret_bf16__offset__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v2, 4.0, v2
+; GFX11-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB19_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_noret_bf16__offset__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX10-NEXT: v_add_f32_e32 v2, 4.0, v2
+; GFX10-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB19_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_noret_bf16__offset__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX90A-NEXT: v_add_f32_e32 v2, 4.0, v2
+; GFX90A-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v3, v3, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v1, s7, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB19_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_noret_bf16__offset__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX908-NEXT: v_add_f32_e32 v2, 4.0, v2
+; GFX908-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v3, v3, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: v_and_or_b32 v2, v1, s7, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB19_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_noret_bf16__offset__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX8-NEXT: v_add_f32_e32 v2, 4.0, v2
+; GFX8-NEXT: v_bfe_u32 v4, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v2
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
+; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc
+; GFX8-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB19_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_noret_bf16__offset__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX7-NEXT: v_add_f32_e32 v2, 4.0, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB19_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_noret_bf16__offset__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffe, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX6-NEXT: v_add_f32_e32 v2, 4.0, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB19_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(3) %ptr, i32 32767
+ %unused = atomicrmw fadd ptr addrspace(3) %gep, bfloat 4.0 seq_cst, align 4
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x half>
+; --------------------------------------------------------------------
+
+define <2 x half> @local_atomic_fadd_ret_v2f16(ptr addrspace(3) %ptr, <2 x half> %val) {
+; GFX12-LABEL: local_atomic_fadd_ret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v2, v3, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB20_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_ret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: v_pk_add_f16 v2, v3, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB20_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_ret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v2, v3, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB20_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_ret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: v_pk_add_f16 v2, v3, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB20_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_ret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: v_pk_add_f16 v2, v3, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB20_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_ret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: v_pk_add_f16 v2, v3, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB20_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_ret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v2, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: v_add_f16_sdwa v2, v3, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v4, v3, v1
+; GFX8-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB20_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_ret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v1
+; GFX7-NEXT: v_add_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v5
+; GFX7-NEXT: v_or_b32_e32 v1, v6, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB20_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_ret_v2f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_or_b32_e32 v7, v2, v1
+; GFX6-NEXT: v_add_f32_e32 v5, v5, v3
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v5
+; GFX6-NEXT: v_or_b32_e32 v1, v6, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB20_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v0, v2
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fadd ptr addrspace(3) %ptr, <2 x half> %val seq_cst
+ ret <2 x half> %result
+}
+
+define <2 x half> @local_atomic_fadd_ret_v2f16__offset(ptr addrspace(3) %ptr, <2 x half> %val) {
+; GFX12-LABEL: local_atomic_fadd_ret_v2f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v2, v3, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB21_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_ret_v2f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: v_pk_add_f16 v2, v3, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB21_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_ret_v2f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v2, v3, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB21_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_ret_v2f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: v_pk_add_f16 v2, v3, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB21_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_ret_v2f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: v_pk_add_f16 v2, v3, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB21_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_ret_v2f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: v_pk_add_f16 v2, v3, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB21_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_ret_v2f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: v_add_f16_sdwa v2, v3, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v4, v3, v1
+; GFX8-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB21_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_ret_v2f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v1
+; GFX7-NEXT: v_add_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v5
+; GFX7-NEXT: v_or_b32_e32 v1, v6, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v1 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB21_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_ret_v2f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v3, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v4, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX6-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v5
+; GFX6-NEXT: v_or_b32_e32 v0, v6, v0
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v3, v7, v0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB21_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fadd ptr addrspace(3) %gep, <2 x half> %val seq_cst
+ ret <2 x half> %result
+}
+
+define void @local_atomic_fadd_noret_v2f16(ptr addrspace(3) %ptr, <2 x half> %val) {
+; GFX12-LABEL: local_atomic_fadd_noret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v3, v2, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX12-NEXT: v_mov_b32_e32 v2, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB22_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_noret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v3, v2, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB22_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_noret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v3, v2, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX11-NEXT: v_mov_b32_e32 v2, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB22_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_noret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v3, v2, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX10-NEXT: v_mov_b32_e32 v2, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB22_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_noret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v3, v2, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB22_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_noret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v3, v2, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v2, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB22_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_noret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v2, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_add_f16_sdwa v3, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v4, v2, v1
+; GFX8-NEXT: v_or_b32_e32 v3, v4, v3
+; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v2, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB22_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_noret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_or_b32_e32 v7, v3, v4
+; GFX7-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_or_b32_e32 v3, v6, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB22_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_noret_v2f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_or_b32_e32 v7, v3, v4
+; GFX6-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_or_b32_e32 v3, v6, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB22_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(3) %ptr, <2 x half> %val seq_cst
+ ret void
+}
+
+define void @local_atomic_fadd_noret_v2f16__offset(ptr addrspace(3) %ptr, <2 x half> %val) {
+; GFX12-LABEL: local_atomic_fadd_noret_v2f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v3, v2, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX12-NEXT: v_mov_b32_e32 v2, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB23_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_noret_v2f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v3, v2, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB23_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_noret_v2f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v3, v2, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX11-NEXT: v_mov_b32_e32 v2, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB23_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_noret_v2f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v3, v2, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX10-NEXT: v_mov_b32_e32 v2, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB23_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_noret_v2f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v3, v2, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB23_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_noret_v2f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v3, v2, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v2, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB23_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_noret_v2f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_add_f16_sdwa v3, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v4, v2, v1
+; GFX8-NEXT: v_or_b32_e32 v3, v4, v3
+; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v2, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB23_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_noret_v2f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_or_b32_e32 v7, v3, v4
+; GFX7-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_or_b32_e32 v3, v6, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v3 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB23_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_noret_v2f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v4, v0
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_or_b32_e32 v7, v3, v4
+; GFX6-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_or_b32_e32 v3, v6, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB23_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fadd ptr addrspace(3) %gep, <2 x half> %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x bfloat>
+; --------------------------------------------------------------------
+
+define <2 x bfloat> @local_atomic_fadd_ret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
+; GFX12-LABEL: local_atomic_fadd_ret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX12-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
+; GFX12-NEXT: v_add_f32_e32 v2, v2, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v6, v8, s0
+; GFX12-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB24_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_ret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX940-NEXT: v_add_f32_e32 v2, v2, v3
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX940-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v2, s4
+; GFX940-NEXT: v_add3_u32 v8, v8, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v5, v2, s5
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB24_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_ret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX11-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
+; GFX11-NEXT: v_add_f32_e32 v2, v2, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v6, v8, s0
+; GFX11-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB24_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_ret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX10-NEXT: v_add_f32_e32 v2, v2, v3
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX10-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v6, v8, s4
+; GFX10-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB24_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_ret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX90A-NEXT: v_add_f32_e32 v2, v2, v3
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX90A-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v2, s8
+; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v5, v2, s9
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB24_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_ret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX908-NEXT: v_add_f32_e32 v2, v2, v3
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX908-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v6, v6, v2, s8
+; GFX908-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v2, v5, v2, s9
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB24_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_ret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v2, v0
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX8-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX8-NEXT: v_add_f32_e32 v2, v2, v3
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX8-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v2, v5, v2, 16
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB24_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_ret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_add_f32_e32 v5, v5, v4
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v3, v3, v5, 16
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v0, v1, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB24_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_ret_v2bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_add_f32_e32 v5, v5, v4
+; GFX6-NEXT: v_alignbit_b32 v1, v1, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v3, v3, v5, 16
+; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v0, v1, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB24_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(3) %ptr, <2 x bfloat> %val seq_cst
+ ret <2 x bfloat> %result
+}
+
+define <2 x bfloat> @local_atomic_fadd_ret_v2bf16__offset(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
+; GFX12-LABEL: local_atomic_fadd_ret_v2bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX12-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
+; GFX12-NEXT: v_add_f32_e32 v2, v2, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v6, v8, s0
+; GFX12-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB25_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_ret_v2bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX940-NEXT: v_add_f32_e32 v2, v2, v3
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX940-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v2, s4
+; GFX940-NEXT: v_add3_u32 v8, v8, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v5, v2, s5
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB25_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_ret_v2bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX11-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
+; GFX11-NEXT: v_add_f32_e32 v2, v2, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v6, v8, s0
+; GFX11-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB25_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_ret_v2bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX10-NEXT: v_add_f32_e32 v2, v2, v3
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX10-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v6, v8, s4
+; GFX10-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB25_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_ret_v2bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX90A-NEXT: v_add_f32_e32 v2, v2, v3
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX90A-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v2, s8
+; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v5, v2, s9
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB25_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_ret_v2bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX908-NEXT: v_add_f32_e32 v2, v2, v3
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX908-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v6, v6, v2, s8
+; GFX908-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v2, v5, v2, s9
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB25_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_ret_v2bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX8-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX8-NEXT: v_add_f32_e32 v2, v2, v3
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX8-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v2, v5, v2, 16
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB25_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_ret_v2bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_add_f32_e32 v5, v5, v4
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v3, v3, v5, 16
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v0, v1, v3 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB25_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_ret_v2bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v0, v4
+; GFX6-NEXT: v_mov_b32_e32 v3, v1
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v0
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_add_f32_e32 v5, v5, v3
+; GFX6-NEXT: v_alignbit_b32 v0, v1, v0, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v1, v1, v5, 16
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v4, v0, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v5
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB25_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fadd ptr addrspace(3) %gep, <2 x bfloat> %val seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @local_atomic_fadd_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
+; GFX12-LABEL: local_atomic_fadd_noret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v3, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_add_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
+; GFX12-NEXT: v_add_f32_e32 v4, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v4, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v4
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB26_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_noret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v3, v0
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX940-NEXT: v_add_f32_e32 v4, v4, v2
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX940-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v4, s4
+; GFX940-NEXT: v_add3_u32 v8, v8, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v4, v4
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[0:1]
+; GFX940-NEXT: v_perm_b32 v4, v5, v4, s5
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB26_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_noret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v3, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_add_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
+; GFX11-NEXT: v_add_f32_e32 v4, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v4, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v4
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB26_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_noret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX10-NEXT: v_add_f32_e32 v4, v4, v2
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX10-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v4, v4
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v4, v6, v8, s4
+; GFX10-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v4
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB26_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_noret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v3, v0
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX90A-NEXT: v_add_f32_e32 v4, v4, v2
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX90A-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v4, s8
+; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v4, v5, v4, s9
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB26_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_noret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v3, v0
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX908-NEXT: v_add_f32_e32 v4, v4, v2
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX908-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v6, v6, v4, s8
+; GFX908-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX908-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v4, v5, v4, s9
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB26_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_noret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v0
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX8-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX8-NEXT: v_add_f32_e32 v4, v4, v2
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v4, v5, v4, 16
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB26_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_noret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v4, v0
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX7-NEXT: v_alignbit_b32 v3, v3, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v5, 16
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB26_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_noret_v2bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v4, v0
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX6-NEXT: v_alignbit_b32 v3, v3, v4, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v5, 16
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB26_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(3) %ptr, <2 x bfloat> %val seq_cst
+ ret void
+}
+
+define void @local_atomic_fadd_noret_v2bf16__ofset(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
+; GFX12-LABEL: local_atomic_fadd_noret_v2bf16__ofset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v3, v0 offset:65532
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_add_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
+; GFX12-NEXT: v_add_f32_e32 v4, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v4, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v4
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB27_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_noret_v2bf16__ofset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX940-NEXT: v_add_f32_e32 v4, v4, v2
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX940-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v4, s4
+; GFX940-NEXT: v_add3_u32 v8, v8, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v4, v4
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[0:1]
+; GFX940-NEXT: v_perm_b32 v4, v5, v4, s5
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB27_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_noret_v2bf16__ofset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v3, v0 offset:65532
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_add_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
+; GFX11-NEXT: v_add_f32_e32 v4, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v4, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v4
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB27_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_noret_v2bf16__ofset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX10-NEXT: v_add_f32_e32 v4, v4, v2
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX10-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v4, v4
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v4, v6, v8, s4
+; GFX10-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v4
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB27_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_noret_v2bf16__ofset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX90A-NEXT: v_add_f32_e32 v4, v4, v2
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX90A-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v4, s8
+; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v4, v5, v4, s9
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB27_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_noret_v2bf16__ofset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX908-NEXT: v_add_f32_e32 v4, v4, v2
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX908-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v6, v6, v4, s8
+; GFX908-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX908-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v4, v5, v4, s9
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB27_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_noret_v2bf16__ofset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX8-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX8-NEXT: v_add_f32_e32 v4, v4, v2
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v4, v5, v4, 16
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB27_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_noret_v2bf16__ofset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v4, v0 offset:65532
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX7-NEXT: v_alignbit_b32 v3, v3, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v5, 16
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB27_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_noret_v2bf16__ofset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v4, v0
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_add_f32_e32 v5, v5, v1
+; GFX6-NEXT: v_alignbit_b32 v3, v3, v4, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v5, 16
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB27_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fadd ptr addrspace(3) %gep, <2 x bfloat> %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; misc
+; --------------------------------------------------------------------
+
+define amdgpu_kernel void @local_ds_fadd(ptr addrspace(1) %out, ptr addrspace(3) %ptrf, i32 %idx) {
+; GFX12-LABEL: local_ds_fadd:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_load_b64 s[2:3], s[0:1], 0x8
+; GFX12-NEXT: s_mov_b32 s5, exec_lo
+; GFX12-NEXT: s_mov_b32 s4, exec_lo
+; GFX12-NEXT: v_mbcnt_lo_u32_b32 v0, s5, 0
+; GFX12-NEXT: ; implicit-def: $vgpr1
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_add_co_i32 s3, s3, 4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_cmpx_eq_u32_e32 0, v0
+; GFX12-NEXT: s_cbranch_execz .LBB28_2
+; GFX12-NEXT: ; %bb.1:
+; GFX12-NEXT: s_bcnt1_i32_b32 s5, s5
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v1, s5
+; GFX12-NEXT: s_lshl_b32 s5, s3, 3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: v_dual_mov_b32 v2, s5 :: v_dual_mul_f32 v1, 0x42280000, v1
+; GFX12-NEXT: ds_add_rtn_f32 v1, v2, v1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: .LBB28_2:
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: s_mov_b32 s6, exec_lo
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_mbcnt_lo_u32_b32 v2, s6, 0
+; GFX12-NEXT: s_mov_b32 s4, exec_lo
+; GFX12-NEXT: v_cmpx_eq_u32_e32 0, v2
+; GFX12-NEXT: s_cbranch_execz .LBB28_4
+; GFX12-NEXT: ; %bb.3:
+; GFX12-NEXT: s_bcnt1_i32_b32 s6, s6
+; GFX12-NEXT: s_lshl_b32 s3, s3, 4
+; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v2, s3 :: v_dual_mul_f32 v1, 0x42280000, v1
+; GFX12-NEXT: ds_add_f32 v2, v1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: .LBB28_4:
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
+; GFX12-NEXT: s_mov_b32 s4, exec_lo
+; GFX12-NEXT: s_brev_b32 s3, 1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
+; GFX12-NEXT: v_add_f32_e32 v1, s5, v0
+; GFX12-NEXT: ; implicit-def: $vgpr0
+; GFX12-NEXT: .LBB28_5: ; %ComputeLoop
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_ctz_i32_b32 s5, s4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: v_readlane_b32 s6, v1, s5
+; GFX12-NEXT: s_lshl_b32 s7, 1, s5
+; GFX12-NEXT: v_writelane_b32 v0, s3, s5
+; GFX12-NEXT: s_and_not1_b32 s4, s4, s7
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_cmp_lg_u32 s4, 0
+; GFX12-NEXT: s_add_f32 s3, s3, s6
+; GFX12-NEXT: s_cbranch_scc1 .LBB28_5
+; GFX12-NEXT: ; %bb.6: ; %ComputeEnd
+; GFX12-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
+; GFX12-NEXT: ; implicit-def: $vgpr1
+; GFX12-NEXT: s_and_saveexec_b32 s4, vcc_lo
+; GFX12-NEXT: s_xor_b32 s4, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execz .LBB28_8
+; GFX12-NEXT: ; %bb.7:
+; GFX12-NEXT: v_dual_mov_b32 v1, s2 :: v_dual_mov_b32 v2, s3
+; GFX12-NEXT: ds_add_rtn_f32 v1, v1, v2
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: .LBB28_8:
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
+; GFX12-NEXT: v_readfirstlane_b32 s2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_add_f32 v0, s2, v0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_store_b32 v1, v0, s[0:1]
+; GFX12-NEXT: s_nop 0
+; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX12-NEXT: s_endpgm
+;
+; GFX940-LABEL: local_ds_fadd:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x8
+; GFX940-NEXT: s_mov_b64 s[4:5], exec
+; GFX940-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
+; GFX940-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: s_add_i32 s3, s3, 4
+; GFX940-NEXT: ; implicit-def: $vgpr1
+; GFX940-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX940-NEXT: s_cbranch_execz .LBB28_2
+; GFX940-NEXT: ; %bb.1:
+; GFX940-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
+; GFX940-NEXT: s_lshl_b32 s8, s3, 3
+; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
+; GFX940-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
+; GFX940-NEXT: v_mov_b32_e32 v2, s8
+; GFX940-NEXT: ds_add_rtn_f32 v1, v2, v1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: .LBB28_2:
+; GFX940-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX940-NEXT: s_mov_b64 s[6:7], exec
+; GFX940-NEXT: v_readfirstlane_b32 s8, v1
+; GFX940-NEXT: v_mbcnt_lo_u32_b32 v1, s6, 0
+; GFX940-NEXT: v_mbcnt_hi_u32_b32 v1, s7, v1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX940-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX940-NEXT: s_cbranch_execz .LBB28_4
+; GFX940-NEXT: ; %bb.3:
+; GFX940-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
+; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
+; GFX940-NEXT: s_lshl_b32 s3, s3, 4
+; GFX940-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
+; GFX940-NEXT: v_mov_b32_e32 v2, s3
+; GFX940-NEXT: ds_add_f32 v2, v1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: .LBB28_4:
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
+; GFX940-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
+; GFX940-NEXT: s_mov_b64 s[4:5], exec
+; GFX940-NEXT: v_add_f32_e32 v2, s8, v0
+; GFX940-NEXT: v_bfrev_b32_e32 v1, 1
+; GFX940-NEXT: ; implicit-def: $vgpr0
+; GFX940-NEXT: .LBB28_5: ; %ComputeLoop
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_ff1_i32_b64 s3, s[4:5]
+; GFX940-NEXT: s_lshl_b64 s[6:7], 1, s3
+; GFX940-NEXT: v_readfirstlane_b32 s8, v1
+; GFX940-NEXT: v_readlane_b32 s9, v2, s3
+; GFX940-NEXT: s_mov_b32 m0, s3
+; GFX940-NEXT: s_andn2_b64 s[4:5], s[4:5], s[6:7]
+; GFX940-NEXT: v_writelane_b32 v0, s8, m0
+; GFX940-NEXT: s_cmp_lg_u64 s[4:5], 0
+; GFX940-NEXT: v_add_f32_e32 v1, s9, v1
+; GFX940-NEXT: s_cbranch_scc1 .LBB28_5
+; GFX940-NEXT: ; %bb.6: ; %ComputeEnd
+; GFX940-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
+; GFX940-NEXT: v_mbcnt_hi_u32_b32 v2, exec_hi, v2
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
+; GFX940-NEXT: ; implicit-def: $vgpr2
+; GFX940-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX940-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execz .LBB28_8
+; GFX940-NEXT: ; %bb.7:
+; GFX940-NEXT: v_mov_b32_e32 v2, s2
+; GFX940-NEXT: ds_add_rtn_f32 v2, v2, v1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: .LBB28_8:
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX940-NEXT: v_readfirstlane_b32 s2, v2
+; GFX940-NEXT: v_mov_b32_e32 v1, 0
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v0, s2, v0
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: global_store_dword v1, v0, s[0:1] sc0 sc1
+; GFX940-NEXT: s_endpgm
+;
+; GFX11-LABEL: local_ds_fadd:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_load_b64 s[2:3], s[0:1], 0x8
+; GFX11-NEXT: s_mov_b32 s5, exec_lo
+; GFX11-NEXT: s_mov_b32 s4, exec_lo
+; GFX11-NEXT: v_mbcnt_lo_u32_b32 v0, s5, 0
+; GFX11-NEXT: ; implicit-def: $vgpr1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_add_i32 s3, s3, 4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_cmpx_eq_u32_e32 0, v0
+; GFX11-NEXT: s_cbranch_execz .LBB28_2
+; GFX11-NEXT: ; %bb.1:
+; GFX11-NEXT: s_bcnt1_i32_b32 s5, s5
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v1, s5
+; GFX11-NEXT: s_lshl_b32 s5, s3, 3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: v_dual_mov_b32 v2, s5 :: v_dual_mul_f32 v1, 0x42280000, v1
+; GFX11-NEXT: ds_add_rtn_f32 v1, v2, v1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: .LBB28_2:
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: s_mov_b32 s6, exec_lo
+; GFX11-NEXT: v_readfirstlane_b32 s4, v1
+; GFX11-NEXT: v_mbcnt_lo_u32_b32 v2, s6, 0
+; GFX11-NEXT: s_mov_b32 s5, exec_lo
+; GFX11-NEXT: v_cmpx_eq_u32_e32 0, v2
+; GFX11-NEXT: s_cbranch_execz .LBB28_4
+; GFX11-NEXT: ; %bb.3:
+; GFX11-NEXT: s_bcnt1_i32_b32 s6, s6
+; GFX11-NEXT: s_lshl_b32 s3, s3, 4
+; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v2, s3 :: v_dual_mul_f32 v1, 0x42280000, v1
+; GFX11-NEXT: ds_add_f32 v2, v1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: .LBB28_4:
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
+; GFX11-NEXT: v_bfrev_b32_e32 v1, 1
+; GFX11-NEXT: s_mov_b32 s3, exec_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
+; GFX11-NEXT: v_add_f32_e32 v2, s4, v0
+; GFX11-NEXT: ; implicit-def: $vgpr0
+; GFX11-NEXT: .LBB28_5: ; %ComputeLoop
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_ctz_i32_b32 s4, s3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readlane_b32 s6, v2, s4
+; GFX11-NEXT: s_lshl_b32 s7, 1, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: s_and_not1_b32 s3, s3, s7
+; GFX11-NEXT: v_writelane_b32 v0, s5, s4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_add_f32_e32 v1, s6, v1
+; GFX11-NEXT: s_cmp_lg_u32 s3, 0
+; GFX11-NEXT: s_cbranch_scc1 .LBB28_5
+; GFX11-NEXT: ; %bb.6: ; %ComputeEnd
+; GFX11-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v2
+; GFX11-NEXT: ; implicit-def: $vgpr2
+; GFX11-NEXT: s_and_saveexec_b32 s3, vcc_lo
+; GFX11-NEXT: s_xor_b32 s3, exec_lo, s3
+; GFX11-NEXT: s_cbranch_execz .LBB28_8
+; GFX11-NEXT: ; %bb.7:
+; GFX11-NEXT: v_mov_b32_e32 v2, s2
+; GFX11-NEXT: ds_add_rtn_f32 v2, v2, v1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: .LBB28_8:
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s3
+; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
+; GFX11-NEXT: v_readfirstlane_b32 s2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_add_f32 v0, s2, v0
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: global_store_b32 v1, v0, s[0:1]
+; GFX11-NEXT: s_nop 0
+; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-NEXT: s_endpgm
+;
+; GFX10-LABEL: local_ds_fadd:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x8
+; GFX10-NEXT: s_mov_b32 s5, exec_lo
+; GFX10-NEXT: ; implicit-def: $vgpr1
+; GFX10-NEXT: v_mbcnt_lo_u32_b32 v0, s5, 0
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: s_add_i32 s3, s3, 4
+; GFX10-NEXT: s_and_saveexec_b32 s4, vcc_lo
+; GFX10-NEXT: s_cbranch_execz .LBB28_2
+; GFX10-NEXT: ; %bb.1:
+; GFX10-NEXT: s_bcnt1_i32_b32 s5, s5
+; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v1, s5
+; GFX10-NEXT: s_lshl_b32 s5, s3, 3
+; GFX10-NEXT: v_mov_b32_e32 v2, s5
+; GFX10-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
+; GFX10-NEXT: ds_add_rtn_f32 v1, v2, v1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: .LBB28_2:
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: v_readfirstlane_b32 s4, v1
+; GFX10-NEXT: v_mbcnt_lo_u32_b32 v2, s6, 0
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v2
+; GFX10-NEXT: s_and_saveexec_b32 s5, vcc_lo
+; GFX10-NEXT: s_cbranch_execz .LBB28_4
+; GFX10-NEXT: ; %bb.3:
+; GFX10-NEXT: s_bcnt1_i32_b32 s6, s6
+; GFX10-NEXT: s_lshl_b32 s3, s3, 4
+; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
+; GFX10-NEXT: v_mov_b32_e32 v2, s3
+; GFX10-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_add_f32 v2, v1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: .LBB28_4:
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
+; GFX10-NEXT: v_bfrev_b32_e32 v1, 1
+; GFX10-NEXT: s_mov_b32 s3, exec_lo
+; GFX10-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
+; GFX10-NEXT: v_add_f32_e32 v2, s4, v0
+; GFX10-NEXT: ; implicit-def: $vgpr0
+; GFX10-NEXT: .LBB28_5: ; %ComputeLoop
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_ff1_i32_b32 s4, s3
+; GFX10-NEXT: v_readfirstlane_b32 s5, v1
+; GFX10-NEXT: v_readlane_b32 s6, v2, s4
+; GFX10-NEXT: s_lshl_b32 s7, 1, s4
+; GFX10-NEXT: s_andn2_b32 s3, s3, s7
+; GFX10-NEXT: v_writelane_b32 v0, s5, s4
+; GFX10-NEXT: v_add_f32_e32 v1, s6, v1
+; GFX10-NEXT: s_cmp_lg_u32 s3, 0
+; GFX10-NEXT: s_cbranch_scc1 .LBB28_5
+; GFX10-NEXT: ; %bb.6: ; %ComputeEnd
+; GFX10-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v2
+; GFX10-NEXT: ; implicit-def: $vgpr2
+; GFX10-NEXT: s_and_saveexec_b32 s3, vcc_lo
+; GFX10-NEXT: s_xor_b32 s3, exec_lo, s3
+; GFX10-NEXT: s_cbranch_execz .LBB28_8
+; GFX10-NEXT: ; %bb.7:
+; GFX10-NEXT: v_mov_b32_e32 v2, s2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_add_rtn_f32 v2, v2, v1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: .LBB28_8:
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s3
+; GFX10-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX10-NEXT: v_readfirstlane_b32 s2, v2
+; GFX10-NEXT: v_mov_b32_e32 v1, 0
+; GFX10-NEXT: v_add_f32_e32 v0, s2, v0
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: global_store_dword v1, v0, s[0:1]
+; GFX10-NEXT: s_endpgm
+;
+; GFX90A-LABEL: local_ds_fadd:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x8
+; GFX90A-NEXT: s_mov_b64 s[4:5], exec
+; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
+; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_add_i32 s3, s3, 4
+; GFX90A-NEXT: ; implicit-def: $vgpr1
+; GFX90A-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX90A-NEXT: s_cbranch_execz .LBB28_2
+; GFX90A-NEXT: ; %bb.1:
+; GFX90A-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
+; GFX90A-NEXT: s_lshl_b32 s8, s3, 3
+; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
+; GFX90A-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
+; GFX90A-NEXT: v_mov_b32_e32 v2, s8
+; GFX90A-NEXT: ds_add_rtn_f32 v1, v2, v1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: .LBB28_2:
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_mov_b64 s[6:7], exec
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v1
+; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v1, s6, 0
+; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v1, s7, v1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX90A-NEXT: s_cbranch_execz .LBB28_4
+; GFX90A-NEXT: ; %bb.3:
+; GFX90A-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
+; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
+; GFX90A-NEXT: s_lshl_b32 s3, s3, 4
+; GFX90A-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
+; GFX90A-NEXT: v_mov_b32_e32 v2, s3
+; GFX90A-NEXT: ds_add_f32 v2, v1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: .LBB28_4:
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
+; GFX90A-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], exec
+; GFX90A-NEXT: v_add_f32_e32 v2, s8, v0
+; GFX90A-NEXT: v_bfrev_b32_e32 v1, 1
+; GFX90A-NEXT: ; implicit-def: $vgpr0
+; GFX90A-NEXT: .LBB28_5: ; %ComputeLoop
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_ff1_i32_b64 s3, s[4:5]
+; GFX90A-NEXT: s_lshl_b64 s[6:7], 1, s3
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v1
+; GFX90A-NEXT: v_readlane_b32 s9, v2, s3
+; GFX90A-NEXT: s_mov_b32 m0, s3
+; GFX90A-NEXT: s_andn2_b64 s[4:5], s[4:5], s[6:7]
+; GFX90A-NEXT: v_writelane_b32 v0, s8, m0
+; GFX90A-NEXT: s_cmp_lg_u64 s[4:5], 0
+; GFX90A-NEXT: v_add_f32_e32 v1, s9, v1
+; GFX90A-NEXT: s_cbranch_scc1 .LBB28_5
+; GFX90A-NEXT: ; %bb.6: ; %ComputeEnd
+; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
+; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v2, exec_hi, v2
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX90A-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execz .LBB28_8
+; GFX90A-NEXT: ; %bb.7:
+; GFX90A-NEXT: v_mov_b32_e32 v2, s2
+; GFX90A-NEXT: ds_add_rtn_f32 v2, v2, v1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: .LBB28_8:
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX90A-NEXT: v_readfirstlane_b32 s2, v2
+; GFX90A-NEXT: v_mov_b32_e32 v1, 0
+; GFX90A-NEXT: v_add_f32_e32 v0, s2, v0
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: global_store_dword v1, v0, s[0:1]
+; GFX90A-NEXT: s_endpgm
+;
+; GFX908-LABEL: local_ds_fadd:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x8
+; GFX908-NEXT: s_mov_b64 s[4:5], exec
+; GFX908-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
+; GFX908-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: s_add_i32 s3, s3, 4
+; GFX908-NEXT: ; implicit-def: $vgpr1
+; GFX908-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX908-NEXT: s_cbranch_execz .LBB28_2
+; GFX908-NEXT: ; %bb.1:
+; GFX908-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
+; GFX908-NEXT: s_lshl_b32 s8, s3, 3
+; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
+; GFX908-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
+; GFX908-NEXT: v_mov_b32_e32 v2, s8
+; GFX908-NEXT: ds_add_rtn_f32 v1, v2, v1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: .LBB28_2:
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_mov_b64 s[6:7], exec
+; GFX908-NEXT: v_readfirstlane_b32 s8, v1
+; GFX908-NEXT: v_mbcnt_lo_u32_b32 v1, s6, 0
+; GFX908-NEXT: v_mbcnt_hi_u32_b32 v1, s7, v1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX908-NEXT: s_cbranch_execz .LBB28_4
+; GFX908-NEXT: ; %bb.3:
+; GFX908-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
+; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
+; GFX908-NEXT: s_lshl_b32 s3, s3, 4
+; GFX908-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
+; GFX908-NEXT: v_mov_b32_e32 v2, s3
+; GFX908-NEXT: ds_add_f32 v2, v1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: .LBB28_4:
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
+; GFX908-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], exec
+; GFX908-NEXT: v_add_f32_e32 v2, s8, v0
+; GFX908-NEXT: v_bfrev_b32_e32 v1, 1
+; GFX908-NEXT: ; implicit-def: $vgpr0
+; GFX908-NEXT: .LBB28_5: ; %ComputeLoop
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_ff1_i32_b64 s3, s[4:5]
+; GFX908-NEXT: s_lshl_b64 s[6:7], 1, s3
+; GFX908-NEXT: v_readfirstlane_b32 s8, v1
+; GFX908-NEXT: v_readlane_b32 s9, v2, s3
+; GFX908-NEXT: s_mov_b32 m0, s3
+; GFX908-NEXT: s_andn2_b64 s[4:5], s[4:5], s[6:7]
+; GFX908-NEXT: v_writelane_b32 v0, s8, m0
+; GFX908-NEXT: s_cmp_lg_u64 s[4:5], 0
+; GFX908-NEXT: v_add_f32_e32 v1, s9, v1
+; GFX908-NEXT: s_cbranch_scc1 .LBB28_5
+; GFX908-NEXT: ; %bb.6: ; %ComputeEnd
+; GFX908-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
+; GFX908-NEXT: v_mbcnt_hi_u32_b32 v2, exec_hi, v2
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
+; GFX908-NEXT: ; implicit-def: $vgpr2
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX908-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execz .LBB28_8
+; GFX908-NEXT: ; %bb.7:
+; GFX908-NEXT: v_mov_b32_e32 v2, s2
+; GFX908-NEXT: ds_add_rtn_f32 v2, v2, v1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: .LBB28_8:
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX908-NEXT: v_readfirstlane_b32 s2, v2
+; GFX908-NEXT: v_mov_b32_e32 v1, 0
+; GFX908-NEXT: v_add_f32_e32 v0, s2, v0
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: global_store_dword v1, v0, s[0:1]
+; GFX908-NEXT: s_endpgm
+;
+; GFX8-LABEL: local_ds_fadd:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x8
+; GFX8-NEXT: s_mov_b64 s[4:5], exec
+; GFX8-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
+; GFX8-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_add_i32 s3, s3, 4
+; GFX8-NEXT: ; implicit-def: $vgpr1
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX8-NEXT: s_cbranch_execz .LBB28_2
+; GFX8-NEXT: ; %bb.1:
+; GFX8-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
+; GFX8-NEXT: s_lshl_b32 s8, s3, 3
+; GFX8-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
+; GFX8-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
+; GFX8-NEXT: v_mov_b32_e32 v2, s8
+; GFX8-NEXT: ds_add_rtn_f32 v1, v2, v1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: .LBB28_2:
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_mov_b64 s[6:7], exec
+; GFX8-NEXT: v_readfirstlane_b32 s8, v1
+; GFX8-NEXT: v_mbcnt_lo_u32_b32 v1, s6, 0
+; GFX8-NEXT: v_mbcnt_hi_u32_b32 v1, s7, v1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX8-NEXT: s_cbranch_execz .LBB28_4
+; GFX8-NEXT: ; %bb.3:
+; GFX8-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
+; GFX8-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
+; GFX8-NEXT: s_lshl_b32 s3, s3, 4
+; GFX8-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
+; GFX8-NEXT: v_mov_b32_e32 v2, s3
+; GFX8-NEXT: ds_add_f32 v2, v1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: .LBB28_4:
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], exec
+; GFX8-NEXT: v_add_f32_e32 v2, s8, v0
+; GFX8-NEXT: v_bfrev_b32_e32 v1, 1
+; GFX8-NEXT: ; implicit-def: $vgpr0
+; GFX8-NEXT: .LBB28_5: ; %ComputeLoop
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_ff1_i32_b64 s3, s[4:5]
+; GFX8-NEXT: s_lshl_b64 s[6:7], 1, s3
+; GFX8-NEXT: v_readfirstlane_b32 s8, v1
+; GFX8-NEXT: v_readlane_b32 s9, v2, s3
+; GFX8-NEXT: s_mov_b32 m0, s3
+; GFX8-NEXT: s_andn2_b64 s[4:5], s[4:5], s[6:7]
+; GFX8-NEXT: v_writelane_b32 v0, s8, m0
+; GFX8-NEXT: s_cmp_lg_u64 s[4:5], 0
+; GFX8-NEXT: v_add_f32_e32 v1, s9, v1
+; GFX8-NEXT: s_cbranch_scc1 .LBB28_5
+; GFX8-NEXT: ; %bb.6: ; %ComputeEnd
+; GFX8-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
+; GFX8-NEXT: v_mbcnt_hi_u32_b32 v2, exec_hi, v2
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
+; GFX8-NEXT: ; implicit-def: $vgpr2
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX8-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execz .LBB28_8
+; GFX8-NEXT: ; %bb.7:
+; GFX8-NEXT: v_mov_b32_e32 v2, s2
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_add_rtn_f32 v2, v2, v1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: .LBB28_8:
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX8-NEXT: v_readfirstlane_b32 s2, v2
+; GFX8-NEXT: v_add_f32_e32 v2, s2, v0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v0, s0
+; GFX8-NEXT: v_mov_b32_e32 v1, s1
+; GFX8-NEXT: flat_store_dword v[0:1], v2
+; GFX8-NEXT: s_endpgm
+;
+; GFX7-LABEL: local_ds_fadd:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x2
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: v_mbcnt_lo_u32_b32_e64 v0, s6, 0
+; GFX7-NEXT: v_mbcnt_hi_u32_b32_e32 v0, s7, v0
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: s_add_i32 s3, s3, 4
+; GFX7-NEXT: ; implicit-def: $vgpr1
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX7-NEXT: s_cbranch_execz .LBB28_4
+; GFX7-NEXT: ; %bb.1:
+; GFX7-NEXT: s_lshl_b32 s8, s3, 3
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: ds_read_b32 v1, v2
+; GFX7-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
+; GFX7-NEXT: v_cvt_f32_ubyte0_e32 v3, s6
+; GFX7-NEXT: v_mul_f32_e32 v3, 0x42280000, v3
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: .LBB28_2: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_add_f32_e32 v1, v4, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v2, v4, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v4
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB28_2
+; GFX7-NEXT: ; %bb.3: ; %Flow18
+; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX7-NEXT: .LBB28_4: ; %Flow19
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: v_readfirstlane_b32 s8, v1
+; GFX7-NEXT: v_mbcnt_lo_u32_b32_e64 v1, s6, 0
+; GFX7-NEXT: v_mbcnt_hi_u32_b32_e32 v1, s7, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX7-NEXT: s_cbranch_execz .LBB28_7
+; GFX7-NEXT: ; %bb.5:
+; GFX7-NEXT: s_lshl_b32 s3, s3, 4
+; GFX7-NEXT: v_mov_b32_e32 v1, s3
+; GFX7-NEXT: ds_read_b32 v3, v1
+; GFX7-NEXT: s_bcnt1_i32_b64 s3, s[6:7]
+; GFX7-NEXT: v_cvt_f32_ubyte0_e32 v2, s3
+; GFX7-NEXT: v_mul_f32_e32 v2, 0x42280000, v2
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: .LBB28_6: ; %atomicrmw.start2
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v4, v3, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB28_6
+; GFX7-NEXT: .LBB28_7: ; %Flow17
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v2, s2
+; GFX7-NEXT: ds_read_b32 v1, v2
+; GFX7-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
+; GFX7-NEXT: v_add_f32_e32 v0, s8, v0
+; GFX7-NEXT: s_mov_b64 s[2:3], 0
+; GFX7-NEXT: .LBB28_8: ; %atomicrmw.start8
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v1
+; GFX7-NEXT: v_add_f32_e32 v1, v3, v0
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v2, v3, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v3
+; GFX7-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX7-NEXT: s_cbranch_execnz .LBB28_8
+; GFX7-NEXT: ; %bb.9: ; %atomicrmw.end7
+; GFX7-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX7-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX7-NEXT: s_mov_b32 s3, 0xf000
+; GFX7-NEXT: s_mov_b32 s2, -1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: buffer_store_dword v1, off, s[0:3], 0
+; GFX7-NEXT: s_endpgm
+;
+; GFX6-LABEL: local_ds_fadd:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x2
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: v_mbcnt_lo_u32_b32_e64 v0, s6, 0
+; GFX6-NEXT: v_mbcnt_hi_u32_b32_e32 v0, s7, v0
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: s_add_i32 s3, s3, 4
+; GFX6-NEXT: ; implicit-def: $vgpr1
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX6-NEXT: s_cbranch_execz .LBB28_4
+; GFX6-NEXT: ; %bb.1:
+; GFX6-NEXT: s_lshl_b32 s8, s3, 3
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: ds_read_b32 v1, v2
+; GFX6-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
+; GFX6-NEXT: v_cvt_f32_ubyte0_e32 v3, s6
+; GFX6-NEXT: v_mul_f32_e32 v3, 0x42280000, v3
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: .LBB28_2: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_add_f32_e32 v1, v4, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v2, v4, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v4
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB28_2
+; GFX6-NEXT: ; %bb.3: ; %Flow16
+; GFX6-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX6-NEXT: .LBB28_4: ; %Flow17
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: v_readfirstlane_b32 s8, v1
+; GFX6-NEXT: v_mbcnt_lo_u32_b32_e64 v1, s6, 0
+; GFX6-NEXT: v_mbcnt_hi_u32_b32_e32 v1, s7, v1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX6-NEXT: s_cbranch_execz .LBB28_7
+; GFX6-NEXT: ; %bb.5:
+; GFX6-NEXT: s_lshl_b32 s3, s3, 4
+; GFX6-NEXT: v_mov_b32_e32 v1, s3
+; GFX6-NEXT: ds_read_b32 v3, v1
+; GFX6-NEXT: s_bcnt1_i32_b64 s3, s[6:7]
+; GFX6-NEXT: v_cvt_f32_ubyte0_e32 v2, s3
+; GFX6-NEXT: v_mul_f32_e32 v2, 0x42280000, v2
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: .LBB28_6: ; %atomicrmw.start2
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v4, v3, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB28_6
+; GFX6-NEXT: .LBB28_7: ; %Flow15
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v2, s2
+; GFX6-NEXT: ds_read_b32 v1, v2
+; GFX6-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
+; GFX6-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
+; GFX6-NEXT: v_add_f32_e32 v0, s8, v0
+; GFX6-NEXT: s_mov_b64 s[2:3], 0
+; GFX6-NEXT: .LBB28_8: ; %atomicrmw.start8
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v3, v1
+; GFX6-NEXT: v_add_f32_e32 v1, v3, v0
+; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v2, v3, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v3
+; GFX6-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX6-NEXT: s_cbranch_execnz .LBB28_8
+; GFX6-NEXT: ; %bb.9: ; %atomicrmw.end7
+; GFX6-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX6-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX6-NEXT: s_mov_b32 s3, 0xf000
+; GFX6-NEXT: s_mov_b32 s2, -1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: buffer_store_dword v1, off, s[0:3], 0
+; GFX6-NEXT: s_endpgm
+ %idx.add = add nuw i32 %idx, 4
+ %shl0 = shl i32 %idx.add, 3
+ %shl1 = shl i32 %idx.add, 4
+ %ptr0 = inttoptr i32 %shl0 to ptr addrspace(3)
+ %ptr1 = inttoptr i32 %shl1 to ptr addrspace(3)
+ %a1 = atomicrmw fadd ptr addrspace(3) %ptr0, float 4.2e+1 seq_cst
+ %a2 = atomicrmw fadd ptr addrspace(3) %ptr1, float 4.2e+1 seq_cst
+ %a3 = atomicrmw fadd ptr addrspace(3) %ptrf, float %a1 seq_cst
+ store float %a3, ptr addrspace(1) %out
+ ret void
+}
+
+define amdgpu_kernel void @local_ds_fadd_one_as(ptr addrspace(1) %out, ptr addrspace(3) %ptrf, i32 %idx) {
+; GFX12-LABEL: local_ds_fadd_one_as:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_load_b64 s[2:3], s[0:1], 0x8
+; GFX12-NEXT: s_mov_b32 s5, exec_lo
+; GFX12-NEXT: s_mov_b32 s4, exec_lo
+; GFX12-NEXT: v_mbcnt_lo_u32_b32 v0, s5, 0
+; GFX12-NEXT: ; implicit-def: $vgpr1
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_add_co_i32 s3, s3, 4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_cmpx_eq_u32_e32 0, v0
+; GFX12-NEXT: s_cbranch_execz .LBB29_2
+; GFX12-NEXT: ; %bb.1:
+; GFX12-NEXT: s_bcnt1_i32_b32 s5, s5
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v1, s5
+; GFX12-NEXT: s_lshl_b32 s5, s3, 3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: v_dual_mov_b32 v2, s5 :: v_dual_mul_f32 v1, 0x42280000, v1
+; GFX12-NEXT: ds_add_rtn_f32 v1, v2, v1
+; GFX12-NEXT: .LBB29_2:
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: s_mov_b32 s6, exec_lo
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_mbcnt_lo_u32_b32 v2, s6, 0
+; GFX12-NEXT: s_mov_b32 s4, exec_lo
+; GFX12-NEXT: v_cmpx_eq_u32_e32 0, v2
+; GFX12-NEXT: s_cbranch_execz .LBB29_4
+; GFX12-NEXT: ; %bb.3:
+; GFX12-NEXT: s_bcnt1_i32_b32 s6, s6
+; GFX12-NEXT: s_lshl_b32 s3, s3, 4
+; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v2, s3 :: v_dual_mul_f32 v1, 0x42280000, v1
+; GFX12-NEXT: ds_add_f32 v2, v1
+; GFX12-NEXT: .LBB29_4:
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
+; GFX12-NEXT: s_mov_b32 s4, exec_lo
+; GFX12-NEXT: s_brev_b32 s3, 1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
+; GFX12-NEXT: v_add_f32_e32 v1, s5, v0
+; GFX12-NEXT: ; implicit-def: $vgpr0
+; GFX12-NEXT: .LBB29_5: ; %ComputeLoop
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_ctz_i32_b32 s5, s4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: v_readlane_b32 s6, v1, s5
+; GFX12-NEXT: s_lshl_b32 s7, 1, s5
+; GFX12-NEXT: v_writelane_b32 v0, s3, s5
+; GFX12-NEXT: s_and_not1_b32 s4, s4, s7
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_cmp_lg_u32 s4, 0
+; GFX12-NEXT: s_add_f32 s3, s3, s6
+; GFX12-NEXT: s_cbranch_scc1 .LBB29_5
+; GFX12-NEXT: ; %bb.6: ; %ComputeEnd
+; GFX12-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
+; GFX12-NEXT: ; implicit-def: $vgpr1
+; GFX12-NEXT: s_and_saveexec_b32 s4, vcc_lo
+; GFX12-NEXT: s_xor_b32 s4, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execz .LBB29_8
+; GFX12-NEXT: ; %bb.7:
+; GFX12-NEXT: v_dual_mov_b32 v1, s2 :: v_dual_mov_b32 v2, s3
+; GFX12-NEXT: ds_add_rtn_f32 v1, v1, v2
+; GFX12-NEXT: .LBB29_8:
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_readfirstlane_b32 s2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_add_f32 v0, s2, v0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_store_b32 v1, v0, s[0:1]
+; GFX12-NEXT: s_nop 0
+; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX12-NEXT: s_endpgm
+;
+; GFX940-LABEL: local_ds_fadd_one_as:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x8
+; GFX940-NEXT: s_mov_b64 s[4:5], exec
+; GFX940-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
+; GFX940-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: s_add_i32 s3, s3, 4
+; GFX940-NEXT: ; implicit-def: $vgpr1
+; GFX940-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX940-NEXT: s_cbranch_execz .LBB29_2
+; GFX940-NEXT: ; %bb.1:
+; GFX940-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
+; GFX940-NEXT: s_lshl_b32 s8, s3, 3
+; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
+; GFX940-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
+; GFX940-NEXT: v_mov_b32_e32 v2, s8
+; GFX940-NEXT: ds_add_rtn_f32 v1, v2, v1
+; GFX940-NEXT: .LBB29_2:
+; GFX940-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX940-NEXT: s_mov_b64 s[6:7], exec
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_readfirstlane_b32 s8, v1
+; GFX940-NEXT: v_mbcnt_lo_u32_b32 v1, s6, 0
+; GFX940-NEXT: v_mbcnt_hi_u32_b32 v1, s7, v1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX940-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX940-NEXT: s_cbranch_execz .LBB29_4
+; GFX940-NEXT: ; %bb.3:
+; GFX940-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
+; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
+; GFX940-NEXT: s_lshl_b32 s3, s3, 4
+; GFX940-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
+; GFX940-NEXT: v_mov_b32_e32 v2, s3
+; GFX940-NEXT: ds_add_f32 v2, v1
+; GFX940-NEXT: .LBB29_4:
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
+; GFX940-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
+; GFX940-NEXT: s_mov_b64 s[4:5], exec
+; GFX940-NEXT: v_add_f32_e32 v2, s8, v0
+; GFX940-NEXT: v_bfrev_b32_e32 v1, 1
+; GFX940-NEXT: ; implicit-def: $vgpr0
+; GFX940-NEXT: .LBB29_5: ; %ComputeLoop
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_ff1_i32_b64 s3, s[4:5]
+; GFX940-NEXT: s_lshl_b64 s[6:7], 1, s3
+; GFX940-NEXT: v_readfirstlane_b32 s8, v1
+; GFX940-NEXT: v_readlane_b32 s9, v2, s3
+; GFX940-NEXT: s_mov_b32 m0, s3
+; GFX940-NEXT: s_andn2_b64 s[4:5], s[4:5], s[6:7]
+; GFX940-NEXT: v_writelane_b32 v0, s8, m0
+; GFX940-NEXT: s_cmp_lg_u64 s[4:5], 0
+; GFX940-NEXT: v_add_f32_e32 v1, s9, v1
+; GFX940-NEXT: s_cbranch_scc1 .LBB29_5
+; GFX940-NEXT: ; %bb.6: ; %ComputeEnd
+; GFX940-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
+; GFX940-NEXT: v_mbcnt_hi_u32_b32 v2, exec_hi, v2
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
+; GFX940-NEXT: ; implicit-def: $vgpr2
+; GFX940-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX940-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execz .LBB29_8
+; GFX940-NEXT: ; %bb.7:
+; GFX940-NEXT: v_mov_b32_e32 v2, s2
+; GFX940-NEXT: ds_add_rtn_f32 v2, v2, v1
+; GFX940-NEXT: .LBB29_8:
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_readfirstlane_b32 s2, v2
+; GFX940-NEXT: v_mov_b32_e32 v1, 0
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v0, s2, v0
+; GFX940-NEXT: global_store_dword v1, v0, s[0:1] sc0 sc1
+; GFX940-NEXT: s_endpgm
+;
+; GFX11-LABEL: local_ds_fadd_one_as:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_load_b64 s[2:3], s[0:1], 0x8
+; GFX11-NEXT: s_mov_b32 s5, exec_lo
+; GFX11-NEXT: s_mov_b32 s4, exec_lo
+; GFX11-NEXT: v_mbcnt_lo_u32_b32 v0, s5, 0
+; GFX11-NEXT: ; implicit-def: $vgpr1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_add_i32 s3, s3, 4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_cmpx_eq_u32_e32 0, v0
+; GFX11-NEXT: s_cbranch_execz .LBB29_2
+; GFX11-NEXT: ; %bb.1:
+; GFX11-NEXT: s_bcnt1_i32_b32 s5, s5
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v1, s5
+; GFX11-NEXT: s_lshl_b32 s5, s3, 3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: v_dual_mov_b32 v2, s5 :: v_dual_mul_f32 v1, 0x42280000, v1
+; GFX11-NEXT: ds_add_rtn_f32 v1, v2, v1
+; GFX11-NEXT: .LBB29_2:
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: s_mov_b32 s6, exec_lo
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_readfirstlane_b32 s4, v1
+; GFX11-NEXT: v_mbcnt_lo_u32_b32 v2, s6, 0
+; GFX11-NEXT: s_mov_b32 s5, exec_lo
+; GFX11-NEXT: v_cmpx_eq_u32_e32 0, v2
+; GFX11-NEXT: s_cbranch_execz .LBB29_4
+; GFX11-NEXT: ; %bb.3:
+; GFX11-NEXT: s_bcnt1_i32_b32 s6, s6
+; GFX11-NEXT: s_lshl_b32 s3, s3, 4
+; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v2, s3 :: v_dual_mul_f32 v1, 0x42280000, v1
+; GFX11-NEXT: ds_add_f32 v2, v1
+; GFX11-NEXT: .LBB29_4:
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
+; GFX11-NEXT: v_bfrev_b32_e32 v1, 1
+; GFX11-NEXT: s_mov_b32 s3, exec_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
+; GFX11-NEXT: v_add_f32_e32 v2, s4, v0
+; GFX11-NEXT: ; implicit-def: $vgpr0
+; GFX11-NEXT: .LBB29_5: ; %ComputeLoop
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_ctz_i32_b32 s4, s3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readlane_b32 s6, v2, s4
+; GFX11-NEXT: s_lshl_b32 s7, 1, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: s_and_not1_b32 s3, s3, s7
+; GFX11-NEXT: v_writelane_b32 v0, s5, s4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_add_f32_e32 v1, s6, v1
+; GFX11-NEXT: s_cmp_lg_u32 s3, 0
+; GFX11-NEXT: s_cbranch_scc1 .LBB29_5
+; GFX11-NEXT: ; %bb.6: ; %ComputeEnd
+; GFX11-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v2
+; GFX11-NEXT: ; implicit-def: $vgpr2
+; GFX11-NEXT: s_and_saveexec_b32 s3, vcc_lo
+; GFX11-NEXT: s_xor_b32 s3, exec_lo, s3
+; GFX11-NEXT: s_cbranch_execz .LBB29_8
+; GFX11-NEXT: ; %bb.7:
+; GFX11-NEXT: v_mov_b32_e32 v2, s2
+; GFX11-NEXT: ds_add_rtn_f32 v2, v2, v1
+; GFX11-NEXT: .LBB29_8:
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s3
+; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_readfirstlane_b32 s2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_add_f32 v0, s2, v0
+; GFX11-NEXT: global_store_b32 v1, v0, s[0:1]
+; GFX11-NEXT: s_nop 0
+; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-NEXT: s_endpgm
+;
+; GFX10-LABEL: local_ds_fadd_one_as:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x8
+; GFX10-NEXT: s_mov_b32 s5, exec_lo
+; GFX10-NEXT: ; implicit-def: $vgpr1
+; GFX10-NEXT: v_mbcnt_lo_u32_b32 v0, s5, 0
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: s_add_i32 s3, s3, 4
+; GFX10-NEXT: s_and_saveexec_b32 s4, vcc_lo
+; GFX10-NEXT: s_cbranch_execz .LBB29_2
+; GFX10-NEXT: ; %bb.1:
+; GFX10-NEXT: s_bcnt1_i32_b32 s5, s5
+; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v1, s5
+; GFX10-NEXT: s_lshl_b32 s5, s3, 3
+; GFX10-NEXT: v_mov_b32_e32 v2, s5
+; GFX10-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
+; GFX10-NEXT: ds_add_rtn_f32 v1, v2, v1
+; GFX10-NEXT: .LBB29_2:
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_readfirstlane_b32 s4, v1
+; GFX10-NEXT: v_mbcnt_lo_u32_b32 v2, s6, 0
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v2
+; GFX10-NEXT: s_and_saveexec_b32 s5, vcc_lo
+; GFX10-NEXT: s_cbranch_execz .LBB29_4
+; GFX10-NEXT: ; %bb.3:
+; GFX10-NEXT: s_bcnt1_i32_b32 s6, s6
+; GFX10-NEXT: s_lshl_b32 s3, s3, 4
+; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
+; GFX10-NEXT: v_mov_b32_e32 v2, s3
+; GFX10-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
+; GFX10-NEXT: ds_add_f32 v2, v1
+; GFX10-NEXT: .LBB29_4:
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
+; GFX10-NEXT: v_bfrev_b32_e32 v1, 1
+; GFX10-NEXT: s_mov_b32 s3, exec_lo
+; GFX10-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
+; GFX10-NEXT: v_add_f32_e32 v2, s4, v0
+; GFX10-NEXT: ; implicit-def: $vgpr0
+; GFX10-NEXT: .LBB29_5: ; %ComputeLoop
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_ff1_i32_b32 s4, s3
+; GFX10-NEXT: v_readfirstlane_b32 s5, v1
+; GFX10-NEXT: v_readlane_b32 s6, v2, s4
+; GFX10-NEXT: s_lshl_b32 s7, 1, s4
+; GFX10-NEXT: s_andn2_b32 s3, s3, s7
+; GFX10-NEXT: v_writelane_b32 v0, s5, s4
+; GFX10-NEXT: v_add_f32_e32 v1, s6, v1
+; GFX10-NEXT: s_cmp_lg_u32 s3, 0
+; GFX10-NEXT: s_cbranch_scc1 .LBB29_5
+; GFX10-NEXT: ; %bb.6: ; %ComputeEnd
+; GFX10-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v2
+; GFX10-NEXT: ; implicit-def: $vgpr2
+; GFX10-NEXT: s_and_saveexec_b32 s3, vcc_lo
+; GFX10-NEXT: s_xor_b32 s3, exec_lo, s3
+; GFX10-NEXT: s_cbranch_execz .LBB29_8
+; GFX10-NEXT: ; %bb.7:
+; GFX10-NEXT: v_mov_b32_e32 v2, s2
+; GFX10-NEXT: ds_add_rtn_f32 v2, v2, v1
+; GFX10-NEXT: .LBB29_8:
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s3
+; GFX10-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_readfirstlane_b32 s2, v2
+; GFX10-NEXT: v_mov_b32_e32 v1, 0
+; GFX10-NEXT: v_add_f32_e32 v0, s2, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_store_dword v1, v0, s[0:1]
+; GFX10-NEXT: s_endpgm
+;
+; GFX90A-LABEL: local_ds_fadd_one_as:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x8
+; GFX90A-NEXT: s_mov_b64 s[4:5], exec
+; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
+; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_add_i32 s3, s3, 4
+; GFX90A-NEXT: ; implicit-def: $vgpr1
+; GFX90A-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX90A-NEXT: s_cbranch_execz .LBB29_2
+; GFX90A-NEXT: ; %bb.1:
+; GFX90A-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
+; GFX90A-NEXT: s_lshl_b32 s8, s3, 3
+; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
+; GFX90A-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
+; GFX90A-NEXT: v_mov_b32_e32 v2, s8
+; GFX90A-NEXT: ds_add_rtn_f32 v1, v2, v1
+; GFX90A-NEXT: .LBB29_2:
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_mov_b64 s[6:7], exec
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v1
+; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v1, s6, 0
+; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v1, s7, v1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX90A-NEXT: s_cbranch_execz .LBB29_4
+; GFX90A-NEXT: ; %bb.3:
+; GFX90A-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
+; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
+; GFX90A-NEXT: s_lshl_b32 s3, s3, 4
+; GFX90A-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
+; GFX90A-NEXT: v_mov_b32_e32 v2, s3
+; GFX90A-NEXT: ds_add_f32 v2, v1
+; GFX90A-NEXT: .LBB29_4:
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
+; GFX90A-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], exec
+; GFX90A-NEXT: v_add_f32_e32 v2, s8, v0
+; GFX90A-NEXT: v_bfrev_b32_e32 v1, 1
+; GFX90A-NEXT: ; implicit-def: $vgpr0
+; GFX90A-NEXT: .LBB29_5: ; %ComputeLoop
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_ff1_i32_b64 s3, s[4:5]
+; GFX90A-NEXT: s_lshl_b64 s[6:7], 1, s3
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v1
+; GFX90A-NEXT: v_readlane_b32 s9, v2, s3
+; GFX90A-NEXT: s_mov_b32 m0, s3
+; GFX90A-NEXT: s_andn2_b64 s[4:5], s[4:5], s[6:7]
+; GFX90A-NEXT: v_writelane_b32 v0, s8, m0
+; GFX90A-NEXT: s_cmp_lg_u64 s[4:5], 0
+; GFX90A-NEXT: v_add_f32_e32 v1, s9, v1
+; GFX90A-NEXT: s_cbranch_scc1 .LBB29_5
+; GFX90A-NEXT: ; %bb.6: ; %ComputeEnd
+; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
+; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v2, exec_hi, v2
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX90A-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execz .LBB29_8
+; GFX90A-NEXT: ; %bb.7:
+; GFX90A-NEXT: v_mov_b32_e32 v2, s2
+; GFX90A-NEXT: ds_add_rtn_f32 v2, v2, v1
+; GFX90A-NEXT: .LBB29_8:
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_readfirstlane_b32 s2, v2
+; GFX90A-NEXT: v_mov_b32_e32 v1, 0
+; GFX90A-NEXT: v_add_f32_e32 v0, s2, v0
+; GFX90A-NEXT: global_store_dword v1, v0, s[0:1]
+; GFX90A-NEXT: s_endpgm
+;
+; GFX908-LABEL: local_ds_fadd_one_as:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x8
+; GFX908-NEXT: s_mov_b64 s[4:5], exec
+; GFX908-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
+; GFX908-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: s_add_i32 s3, s3, 4
+; GFX908-NEXT: ; implicit-def: $vgpr1
+; GFX908-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX908-NEXT: s_cbranch_execz .LBB29_2
+; GFX908-NEXT: ; %bb.1:
+; GFX908-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
+; GFX908-NEXT: s_lshl_b32 s8, s3, 3
+; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
+; GFX908-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
+; GFX908-NEXT: v_mov_b32_e32 v2, s8
+; GFX908-NEXT: ds_add_rtn_f32 v1, v2, v1
+; GFX908-NEXT: .LBB29_2:
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_mov_b64 s[6:7], exec
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_readfirstlane_b32 s8, v1
+; GFX908-NEXT: v_mbcnt_lo_u32_b32 v1, s6, 0
+; GFX908-NEXT: v_mbcnt_hi_u32_b32 v1, s7, v1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX908-NEXT: s_cbranch_execz .LBB29_4
+; GFX908-NEXT: ; %bb.3:
+; GFX908-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
+; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
+; GFX908-NEXT: s_lshl_b32 s3, s3, 4
+; GFX908-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
+; GFX908-NEXT: v_mov_b32_e32 v2, s3
+; GFX908-NEXT: ds_add_f32 v2, v1
+; GFX908-NEXT: .LBB29_4:
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
+; GFX908-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], exec
+; GFX908-NEXT: v_add_f32_e32 v2, s8, v0
+; GFX908-NEXT: v_bfrev_b32_e32 v1, 1
+; GFX908-NEXT: ; implicit-def: $vgpr0
+; GFX908-NEXT: .LBB29_5: ; %ComputeLoop
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_ff1_i32_b64 s3, s[4:5]
+; GFX908-NEXT: s_lshl_b64 s[6:7], 1, s3
+; GFX908-NEXT: v_readfirstlane_b32 s8, v1
+; GFX908-NEXT: v_readlane_b32 s9, v2, s3
+; GFX908-NEXT: s_mov_b32 m0, s3
+; GFX908-NEXT: s_andn2_b64 s[4:5], s[4:5], s[6:7]
+; GFX908-NEXT: v_writelane_b32 v0, s8, m0
+; GFX908-NEXT: s_cmp_lg_u64 s[4:5], 0
+; GFX908-NEXT: v_add_f32_e32 v1, s9, v1
+; GFX908-NEXT: s_cbranch_scc1 .LBB29_5
+; GFX908-NEXT: ; %bb.6: ; %ComputeEnd
+; GFX908-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
+; GFX908-NEXT: v_mbcnt_hi_u32_b32 v2, exec_hi, v2
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
+; GFX908-NEXT: ; implicit-def: $vgpr2
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX908-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execz .LBB29_8
+; GFX908-NEXT: ; %bb.7:
+; GFX908-NEXT: v_mov_b32_e32 v2, s2
+; GFX908-NEXT: ds_add_rtn_f32 v2, v2, v1
+; GFX908-NEXT: .LBB29_8:
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_readfirstlane_b32 s2, v2
+; GFX908-NEXT: v_mov_b32_e32 v1, 0
+; GFX908-NEXT: v_add_f32_e32 v0, s2, v0
+; GFX908-NEXT: global_store_dword v1, v0, s[0:1]
+; GFX908-NEXT: s_endpgm
+;
+; GFX8-LABEL: local_ds_fadd_one_as:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x8
+; GFX8-NEXT: s_mov_b64 s[4:5], exec
+; GFX8-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
+; GFX8-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_add_i32 s3, s3, 4
+; GFX8-NEXT: ; implicit-def: $vgpr1
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX8-NEXT: s_cbranch_execz .LBB29_2
+; GFX8-NEXT: ; %bb.1:
+; GFX8-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
+; GFX8-NEXT: s_lshl_b32 s8, s3, 3
+; GFX8-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
+; GFX8-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
+; GFX8-NEXT: v_mov_b32_e32 v2, s8
+; GFX8-NEXT: ds_add_rtn_f32 v1, v2, v1
+; GFX8-NEXT: .LBB29_2:
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_mov_b64 s[6:7], exec
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_readfirstlane_b32 s8, v1
+; GFX8-NEXT: v_mbcnt_lo_u32_b32 v1, s6, 0
+; GFX8-NEXT: v_mbcnt_hi_u32_b32 v1, s7, v1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX8-NEXT: s_cbranch_execz .LBB29_4
+; GFX8-NEXT: ; %bb.3:
+; GFX8-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
+; GFX8-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
+; GFX8-NEXT: s_lshl_b32 s3, s3, 4
+; GFX8-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
+; GFX8-NEXT: v_mov_b32_e32 v2, s3
+; GFX8-NEXT: ds_add_f32 v2, v1
+; GFX8-NEXT: .LBB29_4:
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], exec
+; GFX8-NEXT: v_add_f32_e32 v2, s8, v0
+; GFX8-NEXT: v_bfrev_b32_e32 v1, 1
+; GFX8-NEXT: ; implicit-def: $vgpr0
+; GFX8-NEXT: .LBB29_5: ; %ComputeLoop
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_ff1_i32_b64 s3, s[4:5]
+; GFX8-NEXT: s_lshl_b64 s[6:7], 1, s3
+; GFX8-NEXT: v_readfirstlane_b32 s8, v1
+; GFX8-NEXT: v_readlane_b32 s9, v2, s3
+; GFX8-NEXT: s_mov_b32 m0, s3
+; GFX8-NEXT: s_andn2_b64 s[4:5], s[4:5], s[6:7]
+; GFX8-NEXT: v_writelane_b32 v0, s8, m0
+; GFX8-NEXT: s_cmp_lg_u64 s[4:5], 0
+; GFX8-NEXT: v_add_f32_e32 v1, s9, v1
+; GFX8-NEXT: s_cbranch_scc1 .LBB29_5
+; GFX8-NEXT: ; %bb.6: ; %ComputeEnd
+; GFX8-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
+; GFX8-NEXT: v_mbcnt_hi_u32_b32 v2, exec_hi, v2
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
+; GFX8-NEXT: ; implicit-def: $vgpr2
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX8-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execz .LBB29_8
+; GFX8-NEXT: ; %bb.7:
+; GFX8-NEXT: v_mov_b32_e32 v2, s2
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_add_rtn_f32 v2, v2, v1
+; GFX8-NEXT: .LBB29_8:
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_readfirstlane_b32 s2, v2
+; GFX8-NEXT: v_add_f32_e32 v2, s2, v0
+; GFX8-NEXT: v_mov_b32_e32 v0, s0
+; GFX8-NEXT: v_mov_b32_e32 v1, s1
+; GFX8-NEXT: flat_store_dword v[0:1], v2
+; GFX8-NEXT: s_endpgm
+;
+; GFX7-LABEL: local_ds_fadd_one_as:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x2
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: v_mbcnt_lo_u32_b32_e64 v0, s6, 0
+; GFX7-NEXT: v_mbcnt_hi_u32_b32_e32 v0, s7, v0
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: s_add_i32 s3, s3, 4
+; GFX7-NEXT: ; implicit-def: $vgpr1
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX7-NEXT: s_cbranch_execz .LBB29_4
+; GFX7-NEXT: ; %bb.1:
+; GFX7-NEXT: s_lshl_b32 s8, s3, 3
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: ds_read_b32 v1, v2
+; GFX7-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
+; GFX7-NEXT: v_cvt_f32_ubyte0_e32 v3, s6
+; GFX7-NEXT: v_mul_f32_e32 v3, 0x42280000, v3
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: .LBB29_2: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_add_f32_e32 v1, v4, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v2, v4, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v4
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB29_2
+; GFX7-NEXT: ; %bb.3: ; %Flow18
+; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX7-NEXT: .LBB29_4: ; %Flow19
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: v_readfirstlane_b32 s8, v1
+; GFX7-NEXT: v_mbcnt_lo_u32_b32_e64 v1, s6, 0
+; GFX7-NEXT: v_mbcnt_hi_u32_b32_e32 v1, s7, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX7-NEXT: s_cbranch_execz .LBB29_7
+; GFX7-NEXT: ; %bb.5:
+; GFX7-NEXT: s_lshl_b32 s3, s3, 4
+; GFX7-NEXT: v_mov_b32_e32 v1, s3
+; GFX7-NEXT: ds_read_b32 v3, v1
+; GFX7-NEXT: s_bcnt1_i32_b64 s3, s[6:7]
+; GFX7-NEXT: v_cvt_f32_ubyte0_e32 v2, s3
+; GFX7-NEXT: v_mul_f32_e32 v2, 0x42280000, v2
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: .LBB29_6: ; %atomicrmw.start2
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v4, v3, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB29_6
+; GFX7-NEXT: .LBB29_7: ; %Flow17
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v2, s2
+; GFX7-NEXT: ds_read_b32 v1, v2
+; GFX7-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
+; GFX7-NEXT: v_add_f32_e32 v0, s8, v0
+; GFX7-NEXT: s_mov_b64 s[2:3], 0
+; GFX7-NEXT: .LBB29_8: ; %atomicrmw.start8
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v1
+; GFX7-NEXT: v_add_f32_e32 v1, v3, v0
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v2, v3, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v3
+; GFX7-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX7-NEXT: s_cbranch_execnz .LBB29_8
+; GFX7-NEXT: ; %bb.9: ; %atomicrmw.end7
+; GFX7-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX7-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX7-NEXT: s_mov_b32 s3, 0xf000
+; GFX7-NEXT: s_mov_b32 s2, -1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: buffer_store_dword v1, off, s[0:3], 0
+; GFX7-NEXT: s_endpgm
+;
+; GFX6-LABEL: local_ds_fadd_one_as:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x2
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: v_mbcnt_lo_u32_b32_e64 v0, s6, 0
+; GFX6-NEXT: v_mbcnt_hi_u32_b32_e32 v0, s7, v0
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: s_add_i32 s3, s3, 4
+; GFX6-NEXT: ; implicit-def: $vgpr1
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX6-NEXT: s_cbranch_execz .LBB29_4
+; GFX6-NEXT: ; %bb.1:
+; GFX6-NEXT: s_lshl_b32 s8, s3, 3
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: ds_read_b32 v1, v2
+; GFX6-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
+; GFX6-NEXT: v_cvt_f32_ubyte0_e32 v3, s6
+; GFX6-NEXT: v_mul_f32_e32 v3, 0x42280000, v3
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: .LBB29_2: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_add_f32_e32 v1, v4, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v2, v4, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v4
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB29_2
+; GFX6-NEXT: ; %bb.3: ; %Flow16
+; GFX6-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX6-NEXT: .LBB29_4: ; %Flow17
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: v_readfirstlane_b32 s8, v1
+; GFX6-NEXT: v_mbcnt_lo_u32_b32_e64 v1, s6, 0
+; GFX6-NEXT: v_mbcnt_hi_u32_b32_e32 v1, s7, v1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX6-NEXT: s_cbranch_execz .LBB29_7
+; GFX6-NEXT: ; %bb.5:
+; GFX6-NEXT: s_lshl_b32 s3, s3, 4
+; GFX6-NEXT: v_mov_b32_e32 v1, s3
+; GFX6-NEXT: ds_read_b32 v3, v1
+; GFX6-NEXT: s_bcnt1_i32_b64 s3, s[6:7]
+; GFX6-NEXT: v_cvt_f32_ubyte0_e32 v2, s3
+; GFX6-NEXT: v_mul_f32_e32 v2, 0x42280000, v2
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: .LBB29_6: ; %atomicrmw.start2
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v4, v3, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB29_6
+; GFX6-NEXT: .LBB29_7: ; %Flow15
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v2, s2
+; GFX6-NEXT: ds_read_b32 v1, v2
+; GFX6-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
+; GFX6-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
+; GFX6-NEXT: v_add_f32_e32 v0, s8, v0
+; GFX6-NEXT: s_mov_b64 s[2:3], 0
+; GFX6-NEXT: .LBB29_8: ; %atomicrmw.start8
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v3, v1
+; GFX6-NEXT: v_add_f32_e32 v1, v3, v0
+; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v2, v3, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v3
+; GFX6-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX6-NEXT: s_cbranch_execnz .LBB29_8
+; GFX6-NEXT: ; %bb.9: ; %atomicrmw.end7
+; GFX6-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX6-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX6-NEXT: s_mov_b32 s3, 0xf000
+; GFX6-NEXT: s_mov_b32 s2, -1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: buffer_store_dword v1, off, s[0:3], 0
+; GFX6-NEXT: s_endpgm
+ %idx.add = add nuw i32 %idx, 4
+ %shl0 = shl i32 %idx.add, 3
+ %shl1 = shl i32 %idx.add, 4
+ %ptr0 = inttoptr i32 %shl0 to ptr addrspace(3)
+ %ptr1 = inttoptr i32 %shl1 to ptr addrspace(3)
+ %a1 = atomicrmw fadd ptr addrspace(3) %ptr0, float 4.2e+1 syncscope("one-as") seq_cst
+ %a2 = atomicrmw fadd ptr addrspace(3) %ptr1, float 4.2e+1 syncscope("one-as") seq_cst
+ %a3 = atomicrmw fadd ptr addrspace(3) %ptrf, float %a1 syncscope("one-as") seq_cst
+ store float %a3, ptr addrspace(1) %out
+ ret void
+}
+
+define float @local_atomic_fadd_ret_f32__amdgpu_ignore_denormal_mode(ptr addrspace(3) %ptr) {
+; GFX12-LABEL: local_atomic_fadd_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_add_rtn_f32 v0, v0, v1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX940-NEXT: ds_add_rtn_f32 v0, v0, v1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_add_rtn_f32 v0, v0, v1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_add_rtn_f32 v0, v0, v1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX90A-NEXT: ds_add_rtn_f32 v0, v0, v1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX908-NEXT: ds_add_rtn_f32 v0, v0, v1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_add_rtn_f32 v0, v0, v1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v1
+; GFX7-NEXT: v_add_f32_e32 v1, 4.0, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB30_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v1
+; GFX6-NEXT: v_add_f32_e32 v1, 4.0, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB30_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v0, v1
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(3) %ptr, float 4.0 seq_cst, !amdgpu.ignore.denormal.mode !0
+ ret float %result
+}
+
+define void @local_atomic_fadd_noret_f32__amdgpu_ignore_denormal_mode(ptr addrspace(3) %ptr) {
+; GFX12-LABEL: local_atomic_fadd_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_add_f32 v0, v1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fadd_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX940-NEXT: ds_add_f32 v0, v1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fadd_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_add_f32 v0, v1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fadd_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_add_f32 v0, v1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fadd_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX90A-NEXT: ds_add_f32 v0, v1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fadd_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX908-NEXT: ds_add_f32 v0, v1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fadd_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, 4.0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_add_f32 v0, v1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fadd_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v2, 4.0, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB31_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fadd_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v2, 4.0, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB31_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(3) %ptr, float 4.0 seq_cst, !amdgpu.ignore.denormal.mode !0
+ ret void
+}
+
+!0 = !{}
diff --git a/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmax.ll b/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmax.ll
new file mode 100644
index 0000000000000..6c227c707fbf1
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmax.ll
@@ -0,0 +1,8456 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 < %s | FileCheck -check-prefix=GFX12 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx940 < %s | FileCheck -check-prefix=GFX940 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx90a < %s | FileCheck -check-prefix=GFX90A %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx908 < %s | FileCheck -check-prefix=GFX908 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tonga < %s | FileCheck -check-prefix=GFX8 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=hawaii < %s | FileCheck -check-prefix=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tahiti < %s | FileCheck -check-prefix=GFX6 %s
+
+; --------------------------------------------------------------------
+; float
+; --------------------------------------------------------------------
+
+define float @local_atomic_fmax_ret_f32(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmax_ret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v1, v2, v2
+; GFX12-NEXT: v_max_num_f32_e32 v1, 4.0, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB0_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_ret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v1
+; GFX940-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX940-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB0_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_ret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX11-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB0_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_ret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v1
+; GFX10-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX10-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB0_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v1
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_ret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v1
+; GFX90A-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX90A-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB0_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_ret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v1
+; GFX908-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX908-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB0_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_ret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v1
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB0_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_ret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB0_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_ret_f32:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v1
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX6-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB0_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v0, v1
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(3) %ptr, float 4.0 seq_cst
+ ret float %result
+}
+
+define float @local_atomic_fmax_ret_f32__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmax_ret_f32__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65532
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v1, v2, v2
+; GFX12-NEXT: v_max_num_f32_e32 v1, 4.0, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB1_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_ret_f32__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v1
+; GFX940-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX940-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB1_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_ret_f32__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65532
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX11-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB1_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_ret_f32__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v1
+; GFX10-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX10-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB1_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v1
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_ret_f32__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v1
+; GFX90A-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX90A-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB1_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_ret_f32__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v1
+; GFX908-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX908-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB1_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_ret_f32__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v1
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB1_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_ret_f32__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB1_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_ret_f32__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v1, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v0, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v0
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v2
+; GFX6-NEXT: v_max_f32_e32 v0, 4.0, v0
+; GFX6-NEXT: ds_cmpst_rtn_b32 v0, v1, v2, v0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v0, v2
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB1_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fmax ptr addrspace(3) %gep, float 4.0 seq_cst
+ ret float %result
+}
+
+define void @local_atomic_fmax_noret_f32(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmax_noret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v1, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, 4.0, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB2_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_noret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX940-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB2_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_noret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB2_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_noret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX10-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB2_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_noret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX90A-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB2_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_noret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX908-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB2_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_noret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX8-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB2_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_noret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX7-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_noret_f32:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX6-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB2_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(3) %ptr, float 4.0 seq_cst
+ ret void
+}
+
+define void @local_atomic_fmax_noret_f32__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmax_noret_f32__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65532
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v1, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, 4.0, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB3_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_noret_f32__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX940-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB3_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_noret_f32__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65532
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB3_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_noret_f32__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX10-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB3_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_noret_f32__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX90A-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB3_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_noret_f32__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX908-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB3_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_noret_f32__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX8-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB3_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_noret_f32__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX7-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB3_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_noret_f32__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX6-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB3_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(3) %ptr, i32 16383
+ %unused = atomicrmw fmax ptr addrspace(3) %gep, float 4.0 seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; double
+; --------------------------------------------------------------------
+
+define double @local_atomic_fmax_ret_f64(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmax_ret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v0
+; GFX12-NEXT: ds_load_b64 v[0:1], v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[0:1], v[3:4], v[3:4]
+; GFX12-NEXT: v_max_num_f64_e32 v[0:1], 4.0, v[0:1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4]
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB4_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_ret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v0
+; GFX940-NEXT: ds_read_b64 v[0:1], v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[0:1]
+; GFX940-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
+; GFX940-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX940-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[4:5], v[0:1]
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB4_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_ret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v0
+; GFX11-NEXT: ds_load_b64 v[0:1], v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX11-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4]
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB4_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_ret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v0
+; GFX10-NEXT: ds_read_b64 v[0:1], v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v1
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX10-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB4_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_ret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v0
+; GFX90A-NEXT: ds_read_b64 v[0:1], v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
+; GFX90A-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX90A-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[4:5], v[0:1]
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB4_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_ret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v0
+; GFX908-NEXT: ds_read_b64 v[0:1], v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v1
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX908-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX908-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB4_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_ret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: v_mov_b32_e32 v2, v0
+; GFX8-NEXT: ds_read_b64 v[0:1], v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v1
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX8-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX8-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB4_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_ret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: v_mov_b32_e32 v2, v0
+; GFX7-NEXT: ds_read_b64 v[0:1], v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX7-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX7-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB4_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_ret_f64:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: v_mov_b32_e32 v2, v0
+; GFX6-NEXT: ds_read_b64 v[0:1], v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX6-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX6-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB4_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(3) %ptr, double 4.0 seq_cst
+ ret double %result
+}
+
+define double @local_atomic_fmax_ret_f64__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmax_ret_f64__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v0
+; GFX12-NEXT: ds_load_b64 v[0:1], v0 offset:65528
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[0:1], v[3:4], v[3:4]
+; GFX12-NEXT: v_max_num_f64_e32 v[0:1], 4.0, v[0:1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4] offset:65528
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB5_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_ret_f64__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v0
+; GFX940-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[0:1]
+; GFX940-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
+; GFX940-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX940-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[4:5], v[0:1] offset:65528
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB5_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_ret_f64__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v0
+; GFX11-NEXT: ds_load_b64 v[0:1], v0 offset:65528
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX11-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4] offset:65528
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB5_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_ret_f64__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v0
+; GFX10-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v1
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX10-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] offset:65528
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB5_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_ret_f64__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v0
+; GFX90A-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
+; GFX90A-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX90A-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[4:5], v[0:1] offset:65528
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB5_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_ret_f64__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v0
+; GFX908-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v1
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX908-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX908-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] offset:65528
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_ret_f64__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: v_mov_b32_e32 v2, v0
+; GFX8-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v1
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX8-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX8-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] offset:65528
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_ret_f64__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: v_mov_b32_e32 v2, v0
+; GFX7-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX7-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX7-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] offset:65528
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_ret_f64__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v2, vcc, 0xfff8, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b64 v[0:1], v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX6-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX6-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB5_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(3) %ptr, i32 8191
+ %result = atomicrmw fmax ptr addrspace(3) %gep, double 4.0 seq_cst
+ ret double %result
+}
+
+define void @local_atomic_fmax_noret_f64(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmax_noret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b64 v[1:2], v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_max_num_f64_e32 v[3:4], v[1:2], v[1:2]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[3:4], 4.0, v[3:4]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2]
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX12-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB6_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_noret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b64 v[2:3], v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX940-NEXT: v_max_f64 v[4:5], v[4:5], 4.0
+; GFX940-NEXT: ds_cmpst_rtn_b64 v[4:5], v0, v[2:3], v[4:5]
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], v[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB6_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_noret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b64 v[1:2], v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2]
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX11-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB6_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_noret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b64 v[1:2], v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX10-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX10-NEXT: v_mov_b32_e32 v1, v3
+; GFX10-NEXT: v_mov_b32_e32 v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB6_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_noret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b64 v[2:3], v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX90A-NEXT: v_max_f64 v[4:5], v[4:5], 4.0
+; GFX90A-NEXT: ds_cmpst_rtn_b64 v[4:5], v0, v[2:3], v[4:5]
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB6_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_noret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b64 v[1:2], v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX908-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
+; GFX908-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX908-NEXT: v_mov_b32_e32 v1, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v2, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB6_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_noret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b64 v[1:2], v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX8-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
+; GFX8-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX8-NEXT: v_mov_b32_e32 v1, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v2, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB6_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_noret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b64 v[1:2], v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX7-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
+; GFX7-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v2, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB6_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_noret_f64:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b64 v[1:2], v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX6-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
+; GFX6-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX6-NEXT: v_mov_b32_e32 v1, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v2, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB6_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(3) %ptr, double 4.0 seq_cst
+ ret void
+}
+
+define void @local_atomic_fmax_noret_f64__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmax_noret_f64__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b64 v[1:2], v0 offset:65528
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_max_num_f64_e32 v[3:4], v[1:2], v[1:2]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[3:4], 4.0, v[3:4]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2] offset:65528
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX12-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB7_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_noret_f64__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b64 v[2:3], v0 offset:65528
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX940-NEXT: v_max_f64 v[4:5], v[4:5], 4.0
+; GFX940-NEXT: ds_cmpst_rtn_b64 v[4:5], v0, v[2:3], v[4:5] offset:65528
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], v[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB7_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_noret_f64__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b64 v[1:2], v0 offset:65528
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2] offset:65528
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX11-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB7_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_noret_f64__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b64 v[1:2], v0 offset:65528
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX10-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX10-NEXT: v_mov_b32_e32 v1, v3
+; GFX10-NEXT: v_mov_b32_e32 v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB7_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_noret_f64__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b64 v[2:3], v0 offset:65528
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX90A-NEXT: v_max_f64 v[4:5], v[4:5], 4.0
+; GFX90A-NEXT: ds_cmpst_rtn_b64 v[4:5], v0, v[2:3], v[4:5] offset:65528
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB7_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_noret_f64__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b64 v[1:2], v0 offset:65528
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX908-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
+; GFX908-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX908-NEXT: v_mov_b32_e32 v1, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v2, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB7_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_noret_f64__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b64 v[1:2], v0 offset:65528
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX8-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
+; GFX8-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX8-NEXT: v_mov_b32_e32 v1, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v2, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB7_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_noret_f64__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b64 v[1:2], v0 offset:65528
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX7-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
+; GFX7-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v2, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB7_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_noret_f64__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v2, vcc, 0xfff8, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b64 v[0:1], v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_max_f64 v[3:4], v[0:1], v[0:1]
+; GFX6-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
+; GFX6-NEXT: ds_cmpst_rtn_b64 v[3:4], v2, v[0:1], v[3:4]
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[0:1]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB7_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(3) %ptr, i32 8191
+ %unused = atomicrmw fmax ptr addrspace(3) %gep, double 4.0 seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; half
+; --------------------------------------------------------------------
+
+define half @local_atomic_fmax_ret_f16(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmax_ret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: ds_load_b32 v3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff
+; GFX12-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX12-NEXT: v_max_num_f16_e32 v3, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v3, 4.0, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB8_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_ret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v2, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v3
+; GFX940-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX940-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX940-NEXT: v_max_f16_e32 v3, 4.0, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX940-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB8_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_ret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: ds_load_b32 v3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff
+; GFX11-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX11-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v3, 4.0, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB8_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_ret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: ds_read_b32 v2, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX10-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_not_b32_e32 v3, v3
+; GFX10-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, v0, v4
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_max_f16_e32 v2, 4.0, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v1, v4, v2
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB8_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_ret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v3
+; GFX90A-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX90A-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX90A-NEXT: v_max_f16_e32 v3, 4.0, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX90A-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB8_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_ret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX908-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX908-NEXT: v_max_f16_e32 v3, 4.0, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_ret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX8-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX8-NEXT: v_max_f16_e32 v3, 4.0, v3
+; GFX8-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_ret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX7-NEXT: v_max_f32_e32 v3, 4.0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB8_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_ret_f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX6-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX6-NEXT: v_max_f32_e32 v3, 4.0, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB8_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(3) %ptr, half 4.0 seq_cst
+ ret half %result
+}
+
+define half @local_atomic_fmax_ret_f16__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmax_ret_f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX12-NEXT: ds_load_b32 v3, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX12-NEXT: v_max_num_f16_e32 v3, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v3, 4.0, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB9_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_ret_f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v3, v1
+; GFX940-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v0, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v3
+; GFX940-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX940-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX940-NEXT: v_max_f16_e32 v3, 4.0, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX940-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB9_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_ret_f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX11-NEXT: ds_load_b32 v3, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX11-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v3, 4.0, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB9_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_ret_f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX10-NEXT: ds_read_b32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v2, v2
+; GFX10-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX10-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX10-NEXT: v_max_f16_e32 v3, 4.0, v3
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB9_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_ret_f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v3, v1
+; GFX90A-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v0, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v3
+; GFX90A-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX90A-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX90A-NEXT: v_max_f16_e32 v3, 4.0, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX90A-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB9_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_ret_f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v3, v1
+; GFX908-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v0, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX908-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX908-NEXT: v_max_f16_e32 v3, 4.0, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB9_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_ret_f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffe, v0
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v0, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX8-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX8-NEXT: v_max_f16_e32 v3, 4.0, v3
+; GFX8-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB9_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_ret_f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX7-NEXT: v_max_f32_e32 v3, 4.0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB9_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_ret_f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX6-NEXT: v_max_f32_e32 v3, 4.0, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB9_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(3) %ptr, i32 32767
+ %result = atomicrmw fmax ptr addrspace(3) %gep, half 4.0 seq_cst
+ ret half %result
+}
+
+define void @local_atomic_fmax_noret_f16(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmax_noret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: ds_load_b32 v2, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX12-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_not_b32_e32 v3, v3
+; GFX12-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX12-NEXT: v_max_num_f16_e32 v4, v4, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v4, 4.0, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX12-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX12-NEXT: v_mov_b32_e32 v2, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB10_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_noret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v2, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX940-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX940-NEXT: v_max_f16_e32 v4, 4.0, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX940-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB10_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_noret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: ds_load_b32 v2, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX11-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_not_b32_e32 v3, v3
+; GFX11-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX11-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v4, 4.0, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX11-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX11-NEXT: v_mov_b32_e32 v2, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB10_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_noret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: ds_read_b32 v2, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX10-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_not_b32_e32 v3, v3
+; GFX10-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX10-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX10-NEXT: v_max_f16_e32 v4, 4.0, v4
+; GFX10-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v1, v2, v4
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX10-NEXT: v_mov_b32_e32 v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB10_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_noret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX90A-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX90A-NEXT: v_max_f16_e32 v4, 4.0, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB10_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_noret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX908-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX908-NEXT: v_max_f16_e32 v4, 4.0, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB10_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_noret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX8-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX8-NEXT: v_max_f16_e32 v4, 4.0, v4
+; GFX8-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB10_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_noret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX7-NEXT: v_max_f32_e32 v4, 4.0, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB10_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_noret_f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX6-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX6-NEXT: v_max_f32_e32 v4, 4.0, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX6-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB10_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(3) %ptr, half 4.0 seq_cst
+ ret void
+}
+
+define void @local_atomic_fmax_noret_f16__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmax_noret_f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX12-NEXT: ds_load_b32 v3, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v4, v4, v4
+; GFX12-NEXT: v_max_num_f16_e32 v4, 4.0, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB11_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_noret_f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX940-NEXT: ds_read_b32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v1, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX940-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX940-NEXT: v_max_f16_e32 v4, 4.0, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX940-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB11_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_noret_f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX11-NEXT: ds_load_b32 v3, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX11-NEXT: v_max_f16_e32 v4, 4.0, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB11_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_noret_f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX10-NEXT: ds_read_b32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v2, v2
+; GFX10-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX10-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX10-NEXT: v_max_f16_e32 v4, 4.0, v4
+; GFX10-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB11_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_noret_f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX90A-NEXT: ds_read_b32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX90A-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX90A-NEXT: v_max_f16_e32 v4, 4.0, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_noret_f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX908-NEXT: ds_read_b32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX908-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX908-NEXT: v_max_f16_e32 v4, 4.0, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_noret_f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0xfffe, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX8-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX8-NEXT: v_max_f16_e32 v4, 4.0, v4
+; GFX8-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_noret_f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX7-NEXT: v_max_f32_e32 v4, 4.0, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB11_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_noret_f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX6-NEXT: v_max_f32_e32 v4, 4.0, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX6-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB11_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(3) %ptr, i32 32767
+ %unused = atomicrmw fmax ptr addrspace(3) %gep, half 4.0 seq_cst
+ ret void
+}
+
+define half @local_atomic_fmax_ret_f16__offset__align4(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmax_ret_f16__offset__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v1, v2, v2
+; GFX12-NEXT: v_max_num_f16_e32 v1, 4.0, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX12-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB12_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_ret_f16__offset__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v1
+; GFX940-NEXT: v_max_f16_e32 v1, v2, v2
+; GFX940-NEXT: v_max_f16_e32 v1, 4.0, v1
+; GFX940-NEXT: v_and_or_b32 v1, v2, s2, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB12_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_ret_f16__offset__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v1, v2, v2
+; GFX11-NEXT: v_max_f16_e32 v1, 4.0, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB12_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_ret_f16__offset__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v1
+; GFX10-NEXT: v_max_f16_e32 v1, v2, v2
+; GFX10-NEXT: v_max_f16_e32 v1, 4.0, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX10-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB12_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v1
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_ret_f16__offset__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v1
+; GFX90A-NEXT: v_max_f16_e32 v1, v2, v2
+; GFX90A-NEXT: v_max_f16_e32 v1, 4.0, v1
+; GFX90A-NEXT: v_and_or_b32 v1, v2, s6, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB12_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_ret_f16__offset__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v1
+; GFX908-NEXT: v_max_f16_e32 v1, v2, v2
+; GFX908-NEXT: v_max_f16_e32 v1, 4.0, v1
+; GFX908-NEXT: v_and_or_b32 v1, v2, s6, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB12_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_ret_f16__offset__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v1
+; GFX8-NEXT: v_max_f16_e32 v1, v2, v2
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX8-NEXT: v_max_f16_e32 v1, 4.0, v1
+; GFX8-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB12_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_ret_f16__offset__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX7-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB12_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_ret_f16__offset__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffe, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX6-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB12_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(3) %ptr, i32 32767
+ %result = atomicrmw fmax ptr addrspace(3) %gep, half 4.0 seq_cst, align 4
+ ret half %result
+}
+
+define void @local_atomic_fmax_noret_f16__offset__align4(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmax_noret_f16__offset__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_max_num_f16_e32 v2, v1, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v2, 4.0, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB13_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_noret_f16__offset__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_max_f16_e32 v2, v1, v1
+; GFX940-NEXT: v_max_f16_e32 v2, 4.0, v2
+; GFX940-NEXT: v_and_or_b32 v2, v1, s2, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB13_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_noret_f16__offset__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_max_f16_e32 v2, v1, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v2, 4.0, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB13_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_noret_f16__offset__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_max_f16_e32 v2, v1, v1
+; GFX10-NEXT: v_max_f16_e32 v2, 4.0, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB13_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_noret_f16__offset__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_max_f16_e32 v2, v1, v1
+; GFX90A-NEXT: v_max_f16_e32 v2, 4.0, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v1, s6, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB13_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_noret_f16__offset__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_max_f16_e32 v2, v1, v1
+; GFX908-NEXT: v_max_f16_e32 v2, 4.0, v2
+; GFX908-NEXT: v_and_or_b32 v2, v1, s6, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB13_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_noret_f16__offset__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_max_f16_e32 v2, v1, v1
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX8-NEXT: v_max_f16_e32 v2, 4.0, v2
+; GFX8-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB13_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_noret_f16__offset__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX7-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB13_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_noret_f16__offset__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffe, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v1
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX6-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB13_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(3) %ptr, i32 32767
+ %unused = atomicrmw fmax ptr addrspace(3) %gep, half 4.0 seq_cst, align 4
+ ret void
+}
+
+; --------------------------------------------------------------------
+; bfloat
+; --------------------------------------------------------------------
+
+define bfloat @local_atomic_fmax_ret_bf16(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmax_ret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: ds_load_b32 v3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff
+; GFX12-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, 4.0, v3
+; GFX12-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB14_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_ret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v2, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v3
+; GFX940-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_max_f32_e32 v3, 4.0, v3
+; GFX940-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX940-NEXT: v_add3_u32 v5, v5, v3, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB14_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_ret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: ds_load_b32 v3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff
+; GFX11-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, 4.0, v3
+; GFX11-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB14_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_ret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: ds_read_b32 v3, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff
+; GFX10-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_not_b32_e32 v2, v2
+; GFX10-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_max_f32_e32 v3, 4.0, v3
+; GFX10-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX10-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB14_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_ret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v3
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_max_f32_e32 v3, 4.0, v3
+; GFX90A-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX90A-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_ret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_max_f32_e32 v3, 4.0, v3
+; GFX908-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_ret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f32_e32 v3, 4.0, v3
+; GFX8-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v3
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v3, v6, v7, vcc
+; GFX8-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX8-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_ret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v3, 4.0, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_ret_bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX6-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_max_f32_e32 v3, 4.0, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB14_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(3) %ptr, bfloat 4.0 seq_cst
+ ret bfloat %result
+}
+
+define bfloat @local_atomic_fmax_ret_bf16__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmax_ret_bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX12-NEXT: ds_load_b32 v3, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, 4.0, v3
+; GFX12-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB15_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_ret_bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v3, v1
+; GFX940-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v0, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v3
+; GFX940-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_max_f32_e32 v3, 4.0, v3
+; GFX940-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX940-NEXT: v_add3_u32 v5, v5, v3, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB15_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_ret_bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX11-NEXT: ds_load_b32 v3, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, 4.0, v3
+; GFX11-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB15_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_ret_bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX10-NEXT: ds_read_b32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v2, v2
+; GFX10-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_lshrrev_b32_sdwa v3, v1, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_max_f32_e32 v3, 4.0, v3
+; GFX10-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX10-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB15_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_ret_bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v3, v1
+; GFX90A-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v0, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v3
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_max_f32_e32 v3, 4.0, v3
+; GFX90A-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX90A-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB15_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_ret_bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v3, v1
+; GFX908-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v0, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_max_f32_e32 v3, 4.0, v3
+; GFX908-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB15_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_ret_bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffe, v0
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v0, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f32_e32 v3, 4.0, v3
+; GFX8-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v3
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v3, v6, v7, vcc
+; GFX8-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX8-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB15_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_ret_bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v3, 4.0, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB15_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_ret_bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_max_f32_e32 v3, 4.0, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB15_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(3) %ptr, i32 32767
+ %result = atomicrmw fmax ptr addrspace(3) %gep, bfloat 4.0 seq_cst
+ ret bfloat %result
+}
+
+define void @local_atomic_fmax_noret_bf16(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmax_noret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: ds_load_b32 v2, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX12-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_not_b32_e32 v3, v3
+; GFX12-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v4, 4.0, v4
+; GFX12-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX12-NEXT: v_mov_b32_e32 v2, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB16_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_noret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v2, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_max_f32_e32 v4, 4.0, v4
+; GFX940-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX940-NEXT: v_add3_u32 v5, v5, v4, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB16_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_noret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: ds_load_b32 v2, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX11-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_not_b32_e32 v3, v3
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v4, 4.0, v4
+; GFX11-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX11-NEXT: v_mov_b32_e32 v2, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB16_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_noret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: ds_read_b32 v2, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX10-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_not_b32_e32 v3, v3
+; GFX10-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v4, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_max_f32_e32 v4, 4.0, v4
+; GFX10-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX10-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v1, v2, v4
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX10-NEXT: v_mov_b32_e32 v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB16_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_noret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_max_f32_e32 v4, 4.0, v4
+; GFX90A-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX90A-NEXT: v_add3_u32 v5, v5, v4, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB16_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_noret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_max_f32_e32 v4, 4.0, v4
+; GFX908-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX908-NEXT: v_add3_u32 v5, v5, v4, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX908-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB16_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_noret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f32_e32 v4, 4.0, v4
+; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v4, v6, v7, vcc
+; GFX8-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX8-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB16_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_noret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_max_f32_e32 v4, 4.0, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB16_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_noret_bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX6-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_max_f32_e32 v4, 4.0, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX6-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB16_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(3) %ptr, bfloat 4.0 seq_cst
+ ret void
+}
+
+define void @local_atomic_fmax_noret_bf16__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmax_noret_bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX12-NEXT: ds_load_b32 v3, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: v_max_num_f32_e32 v4, 4.0, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX12-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX12-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB17_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_noret_bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX940-NEXT: ds_read_b32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v1, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_max_f32_e32 v4, 4.0, v4
+; GFX940-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX940-NEXT: v_add3_u32 v5, v5, v4, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB17_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_noret_bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX11-NEXT: ds_load_b32 v3, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: v_max_f32_e32 v4, 4.0, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX11-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB17_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_noret_bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX10-NEXT: ds_read_b32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v2, v2
+; GFX10-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_max_f32_e32 v4, 4.0, v4
+; GFX10-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX10-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB17_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_noret_bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX90A-NEXT: ds_read_b32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_max_f32_e32 v4, 4.0, v4
+; GFX90A-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX90A-NEXT: v_add3_u32 v5, v5, v4, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB17_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_noret_bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX908-NEXT: ds_read_b32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_max_f32_e32 v4, 4.0, v4
+; GFX908-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX908-NEXT: v_add3_u32 v5, v5, v4, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX908-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_noret_bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0xfffe, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f32_e32 v4, 4.0, v4
+; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v4, v6, v7, vcc
+; GFX8-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX8-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_noret_bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_max_f32_e32 v4, 4.0, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_noret_bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_max_f32_e32 v4, 4.0, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX6-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB17_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(3) %ptr, i32 32767
+ %unused = atomicrmw fmax ptr addrspace(3) %gep, bfloat 4.0 seq_cst
+ ret void
+}
+
+define bfloat @local_atomic_fmax_ret_bf16__offset__align4(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmax_ret_bf16__offset__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX12-NEXT: v_max_num_f32_e32 v1, 4.0, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX12-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB18_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_ret_bf16__offset__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX940-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX940-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX940-NEXT: v_add3_u32 v3, v3, v1, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX940-NEXT: v_and_or_b32 v1, v2, s3, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB18_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_ret_bf16__offset__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX11-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB18_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_ret_bf16__offset__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v1
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX10-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX10-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX10-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX10-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB18_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v1
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_ret_bf16__offset__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX90A-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX90A-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX90A-NEXT: v_add3_u32 v3, v3, v1, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX90A-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX90A-NEXT: v_and_or_b32 v1, v2, s7, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB18_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_ret_bf16__offset__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX908-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX908-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX908-NEXT: v_add3_u32 v3, v3, v1, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX908-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX908-NEXT: v_and_or_b32 v1, v2, s7, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB18_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_ret_bf16__offset__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX8-NEXT: v_bfe_u32 v4, v1, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v1
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
+; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v1
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v1, v4, v5, vcc
+; GFX8-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB18_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_ret_bf16__offset__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB18_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_ret_bf16__offset__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffe, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB18_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(3) %ptr, i32 32767
+ %result = atomicrmw fmax ptr addrspace(3) %gep, bfloat 4.0 seq_cst, align 4
+ ret bfloat %result
+}
+
+define void @local_atomic_fmax_noret_bf16__offset__align4(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmax_noret_bf16__offset__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, 4.0, v2
+; GFX12-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB19_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_noret_bf16__offset__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX940-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX940-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v3, v3, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: v_and_or_b32 v2, v1, s3, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB19_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_noret_bf16__offset__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX11-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB19_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_noret_bf16__offset__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX10-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX10-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB19_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_noret_bf16__offset__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX90A-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX90A-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v3, v3, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v1, s7, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB19_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_noret_bf16__offset__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX908-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX908-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v3, v3, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: v_and_or_b32 v2, v1, s7, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB19_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_noret_bf16__offset__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX8-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX8-NEXT: v_bfe_u32 v4, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v2
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
+; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc
+; GFX8-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB19_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_noret_bf16__offset__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB19_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_noret_bf16__offset__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffe, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB19_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(3) %ptr, i32 32767
+ %unused = atomicrmw fmax ptr addrspace(3) %gep, bfloat 4.0 seq_cst, align 4
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x half>
+; --------------------------------------------------------------------
+
+define <2 x half> @local_atomic_fmax_ret_v2f16(ptr addrspace(3) %ptr, <2 x half> %val) {
+; GFX12-LABEL: local_atomic_fmax_ret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0
+; GFX12-NEXT: v_pk_max_num_f16 v1, v1, v1
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v2, v3, v3
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB20_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_ret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX940-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB20_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_ret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0
+; GFX11-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB20_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_ret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0
+; GFX10-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX10-NEXT: v_pk_max_f16 v2, v2, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB20_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_ret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX90A-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB20_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_ret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX908-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB20_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_ret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v2, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v3, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v1, v1, v1
+; GFX8-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: v_max_f16_sdwa v2, v4, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v4, v4
+; GFX8-NEXT: v_max_f16_sdwa v2, v2, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v1
+; GFX8-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB20_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_ret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v1
+; GFX7-NEXT: v_max_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v5
+; GFX7-NEXT: v_or_b32_e32 v1, v6, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB20_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_ret_v2f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_or_b32_e32 v7, v2, v1
+; GFX6-NEXT: v_max_f32_e32 v5, v5, v3
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v5
+; GFX6-NEXT: v_or_b32_e32 v1, v6, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB20_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v0, v2
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fmax ptr addrspace(3) %ptr, <2 x half> %val seq_cst
+ ret <2 x half> %result
+}
+
+define <2 x half> @local_atomic_fmax_ret_v2f16__offset(ptr addrspace(3) %ptr, <2 x half> %val) {
+; GFX12-LABEL: local_atomic_fmax_ret_v2f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX12-NEXT: v_pk_max_num_f16 v1, v1, v1
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v2, v3, v3
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB21_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_ret_v2f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX940-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB21_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_ret_v2f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX11-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB21_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_ret_v2f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX10-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX10-NEXT: v_pk_max_f16 v2, v2, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB21_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_ret_v2f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX90A-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB21_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_ret_v2f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX908-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB21_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_ret_v2f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v3, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v1, v1, v1
+; GFX8-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: v_max_f16_sdwa v2, v4, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v4, v4
+; GFX8-NEXT: v_max_f16_sdwa v2, v2, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v1
+; GFX8-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB21_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_ret_v2f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v1
+; GFX7-NEXT: v_max_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v5
+; GFX7-NEXT: v_or_b32_e32 v1, v6, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v1 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB21_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_ret_v2f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v3, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v4, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX6-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v5
+; GFX6-NEXT: v_or_b32_e32 v0, v6, v0
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v3, v7, v0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB21_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fmax ptr addrspace(3) %gep, <2 x half> %val seq_cst
+ ret <2 x half> %result
+}
+
+define void @local_atomic_fmax_noret_v2f16(ptr addrspace(3) %ptr, <2 x half> %val) {
+; GFX12-LABEL: local_atomic_fmax_noret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0
+; GFX12-NEXT: v_pk_max_num_f16 v1, v1, v1
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v3, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v3, v3, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX12-NEXT: v_mov_b32_e32 v2, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB22_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_noret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX940-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v3, v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_max_f16 v3, v3, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB22_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_noret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0
+; GFX11-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v3, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v3, v3, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX11-NEXT: v_mov_b32_e32 v2, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB22_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_noret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0
+; GFX10-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v3, v2, v2
+; GFX10-NEXT: v_pk_max_f16 v3, v3, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX10-NEXT: v_mov_b32_e32 v2, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB22_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_noret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX90A-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v3, v2, v2
+; GFX90A-NEXT: v_pk_max_f16 v3, v3, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB22_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_noret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX908-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v3, v2, v2
+; GFX908-NEXT: v_pk_max_f16 v3, v3, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v2, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB22_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_noret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v2, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v1, v1, v1
+; GFX8-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v4, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v3, v3
+; GFX8-NEXT: v_max_f16_sdwa v4, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v1
+; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB22_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_noret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_or_b32_e32 v7, v3, v4
+; GFX7-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_or_b32_e32 v3, v6, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB22_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_noret_v2f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_or_b32_e32 v7, v3, v4
+; GFX6-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_or_b32_e32 v3, v6, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB22_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(3) %ptr, <2 x half> %val seq_cst
+ ret void
+}
+
+define void @local_atomic_fmax_noret_v2f16__offset(ptr addrspace(3) %ptr, <2 x half> %val) {
+; GFX12-LABEL: local_atomic_fmax_noret_v2f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX12-NEXT: v_pk_max_num_f16 v1, v1, v1
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v3, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v3, v3, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX12-NEXT: v_mov_b32_e32 v2, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB23_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_noret_v2f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX940-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v3, v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_max_f16 v3, v3, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB23_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_noret_v2f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX11-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v3, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v3, v3, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX11-NEXT: v_mov_b32_e32 v2, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB23_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_noret_v2f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX10-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v3, v2, v2
+; GFX10-NEXT: v_pk_max_f16 v3, v3, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX10-NEXT: v_mov_b32_e32 v2, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB23_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_noret_v2f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX90A-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v3, v2, v2
+; GFX90A-NEXT: v_pk_max_f16 v3, v3, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB23_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_noret_v2f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX908-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v3, v2, v2
+; GFX908-NEXT: v_pk_max_f16 v3, v3, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v2, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB23_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_noret_v2f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v2, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v1, v1, v1
+; GFX8-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v4, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v3, v3
+; GFX8-NEXT: v_max_f16_sdwa v4, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v1
+; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB23_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_noret_v2f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_or_b32_e32 v7, v3, v4
+; GFX7-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_or_b32_e32 v3, v6, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v3 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB23_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_noret_v2f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v4, v0
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_or_b32_e32 v7, v3, v4
+; GFX6-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_or_b32_e32 v3, v6, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB23_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fmax ptr addrspace(3) %gep, <2 x half> %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x bfloat>
+; --------------------------------------------------------------------
+
+define <2 x bfloat> @local_atomic_fmax_ret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
+; GFX12-LABEL: local_atomic_fmax_ret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX12-NEXT: v_max_num_f32_e32 v5, v5, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v6, v8, s0
+; GFX12-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB24_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_ret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v3
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX940-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v2, s4
+; GFX940-NEXT: v_add3_u32 v8, v8, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v5, v2, s5
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB24_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_ret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX11-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v6, v8, s0
+; GFX11-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB24_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_ret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v3
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX10-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v6, v8, s4
+; GFX10-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB24_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_ret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v3
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX90A-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v2, s8
+; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v5, v2, s9
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB24_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_ret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v3
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX908-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v6, v6, v2, s8
+; GFX908-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v2, v5, v2, s9
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB24_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_ret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v2, v0
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX8-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v3
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX8-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v2, v5, v2, 16
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB24_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_ret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
+; GFX7-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_alignbit_b32 v3, v3, v6, 16
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v0, v1, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB24_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_ret_v2bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
+; GFX6-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_alignbit_b32 v1, v1, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_alignbit_b32 v3, v3, v6, 16
+; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v0, v1, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB24_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(3) %ptr, <2 x bfloat> %val seq_cst
+ ret <2 x bfloat> %result
+}
+
+define <2 x bfloat> @local_atomic_fmax_ret_v2bf16__offset(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
+; GFX12-LABEL: local_atomic_fmax_ret_v2bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX12-NEXT: v_max_num_f32_e32 v5, v5, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v6, v8, s0
+; GFX12-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB25_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_ret_v2bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v3
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX940-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v2, s4
+; GFX940-NEXT: v_add3_u32 v8, v8, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v5, v2, s5
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB25_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_ret_v2bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX11-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v6, v8, s0
+; GFX11-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB25_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_ret_v2bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v3
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX10-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v6, v8, s4
+; GFX10-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB25_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_ret_v2bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v3
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX90A-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v2, s8
+; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v5, v2, s9
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB25_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_ret_v2bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v3
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX908-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v6, v6, v2, s8
+; GFX908-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v2, v5, v2, s9
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB25_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_ret_v2bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX8-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v3
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX8-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v2, v5, v2, 16
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB25_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_ret_v2bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
+; GFX7-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_alignbit_b32 v3, v3, v6, 16
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v0, v1, v3 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB25_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_ret_v2bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v0, v4
+; GFX6-NEXT: v_mov_b32_e32 v3, v1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX6-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v0
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v3
+; GFX6-NEXT: v_alignbit_b32 v0, v1, v0, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v5
+; GFX6-NEXT: v_alignbit_b32 v1, v1, v6, 16
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v4, v0, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v5
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB25_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fmax ptr addrspace(3) %gep, <2 x bfloat> %val seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @local_atomic_fmax_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
+; GFX12-LABEL: local_atomic_fmax_noret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v3, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_max_num_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
+; GFX12-NEXT: v_max_num_f32_e32 v4, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v4, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v4
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB26_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_noret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v3, v0
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX940-NEXT: v_max_f32_e32 v4, v4, v2
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX940-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v4, s4
+; GFX940-NEXT: v_add3_u32 v8, v8, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v4, v4
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[0:1]
+; GFX940-NEXT: v_perm_b32 v4, v5, v4, s5
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB26_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_noret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v3, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_max_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
+; GFX11-NEXT: v_max_f32_e32 v4, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v4, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v4
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB26_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_noret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX10-NEXT: v_max_f32_e32 v4, v4, v2
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX10-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v4, v4
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v4, v6, v8, s4
+; GFX10-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v4
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB26_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_noret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v3, v0
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX90A-NEXT: v_max_f32_e32 v4, v4, v2
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX90A-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v4, s8
+; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v4, v5, v4, s9
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB26_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_noret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v3, v0
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX908-NEXT: v_max_f32_e32 v4, v4, v2
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX908-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v6, v6, v4, s8
+; GFX908-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX908-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v4, v5, v4, s9
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB26_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_noret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v0
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX8-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX8-NEXT: v_max_f32_e32 v4, v4, v2
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v4, v5, v4, 16
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB26_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_noret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX7-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v1
+; GFX7-NEXT: v_alignbit_b32 v3, v4, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB26_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_noret_v2bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX6-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v1
+; GFX6-NEXT: v_alignbit_b32 v3, v4, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v0, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB26_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(3) %ptr, <2 x bfloat> %val seq_cst
+ ret void
+}
+
+define void @local_atomic_fmax_noret_v2bf16__ofset(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
+; GFX12-LABEL: local_atomic_fmax_noret_v2bf16__ofset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v3, v0 offset:65532
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_max_num_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
+; GFX12-NEXT: v_max_num_f32_e32 v4, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v4, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v4
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB27_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_noret_v2bf16__ofset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX940-NEXT: v_max_f32_e32 v4, v4, v2
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX940-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v4, s4
+; GFX940-NEXT: v_add3_u32 v8, v8, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v4, v4
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[0:1]
+; GFX940-NEXT: v_perm_b32 v4, v5, v4, s5
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB27_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_noret_v2bf16__ofset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v3, v0 offset:65532
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_max_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
+; GFX11-NEXT: v_max_f32_e32 v4, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v4, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v4
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB27_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_noret_v2bf16__ofset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX10-NEXT: v_max_f32_e32 v4, v4, v2
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX10-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v4, v4
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v4, v6, v8, s4
+; GFX10-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v4
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB27_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_noret_v2bf16__ofset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX90A-NEXT: v_max_f32_e32 v4, v4, v2
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX90A-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v4, s8
+; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v4, v5, v4, s9
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB27_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_noret_v2bf16__ofset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX908-NEXT: v_max_f32_e32 v4, v4, v2
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX908-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v6, v6, v4, s8
+; GFX908-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX908-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v4, v5, v4, s9
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB27_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_noret_v2bf16__ofset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX8-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX8-NEXT: v_max_f32_e32 v4, v4, v2
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v1
+; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v4, v5, v4, 16
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB27_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_noret_v2bf16__ofset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX7-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v1
+; GFX7-NEXT: v_alignbit_b32 v3, v4, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v3, v4 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB27_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_noret_v2bf16__ofset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX6-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v1
+; GFX6-NEXT: v_alignbit_b32 v3, v4, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v0, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB27_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fmax ptr addrspace(3) %gep, <2 x bfloat> %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; misc
+; --------------------------------------------------------------------
+
+define float @local_atomic_fmax_ret_f32__amdgpu_ignore_denormal_mode(ptr addrspace(3) %ptr) {
+; GFX12-LABEL: local_atomic_fmax_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v1, v2, v2
+; GFX12-NEXT: v_max_num_f32_e32 v1, 4.0, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB28_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v1
+; GFX940-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX940-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB28_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX11-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB28_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v1
+; GFX10-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX10-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB28_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v1
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v1
+; GFX90A-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX90A-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB28_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v1
+; GFX908-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX908-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB28_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v1
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB28_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB28_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v1
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX6-NEXT: v_max_f32_e32 v1, 4.0, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB28_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v0, v1
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(3) %ptr, float 4.0 seq_cst, !amdgpu.ignore.denormal.mode !0
+ ret float %result
+}
+
+define void @local_atomic_fmax_noret_f32__amdgpu_ignore_denormal_mode(ptr addrspace(3) %ptr) {
+; GFX12-LABEL: local_atomic_fmax_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v1, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, 4.0, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB29_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmax_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX940-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB29_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmax_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB29_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmax_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX10-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB29_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmax_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX90A-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB29_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmax_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX908-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB29_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmax_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX8-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB29_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmax_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX7-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB29_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmax_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX6-NEXT: v_max_f32_e32 v2, 4.0, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB29_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(3) %ptr, float 4.0 seq_cst, !amdgpu.ignore.denormal.mode !0
+ ret void
+}
+
+!0 = !{}
diff --git a/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmin.ll b/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmin.ll
new file mode 100644
index 0000000000000..c4c70d900a480
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmin.ll
@@ -0,0 +1,8456 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 < %s | FileCheck -check-prefix=GFX12 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx940 < %s | FileCheck -check-prefix=GFX940 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx90a < %s | FileCheck -check-prefix=GFX90A %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx908 < %s | FileCheck -check-prefix=GFX908 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tonga < %s | FileCheck -check-prefix=GFX8 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=hawaii < %s | FileCheck -check-prefix=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tahiti < %s | FileCheck -check-prefix=GFX6 %s
+
+; --------------------------------------------------------------------
+; float
+; --------------------------------------------------------------------
+
+define float @local_atomic_fmin_ret_f32(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmin_ret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v1, v2, v2
+; GFX12-NEXT: v_min_num_f32_e32 v1, 4.0, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB0_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_ret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v1
+; GFX940-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX940-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB0_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_ret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX11-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB0_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_ret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v1
+; GFX10-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX10-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB0_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v1
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_ret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v1
+; GFX90A-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX90A-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB0_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_ret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v1
+; GFX908-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX908-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB0_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_ret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v1
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB0_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_ret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB0_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_ret_f32:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v1
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX6-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB0_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v0, v1
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(3) %ptr, float 4.0 seq_cst
+ ret float %result
+}
+
+define float @local_atomic_fmin_ret_f32__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmin_ret_f32__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65532
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v1, v2, v2
+; GFX12-NEXT: v_min_num_f32_e32 v1, 4.0, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB1_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_ret_f32__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v1
+; GFX940-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX940-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB1_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_ret_f32__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65532
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX11-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB1_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_ret_f32__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v1
+; GFX10-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX10-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB1_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v1
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_ret_f32__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v1
+; GFX90A-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX90A-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB1_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_ret_f32__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v1
+; GFX908-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX908-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB1_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_ret_f32__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v1
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB1_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_ret_f32__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB1_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_ret_f32__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v1, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v0, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v0
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v2
+; GFX6-NEXT: v_min_f32_e32 v0, 4.0, v0
+; GFX6-NEXT: ds_cmpst_rtn_b32 v0, v1, v2, v0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v0, v2
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB1_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fmin ptr addrspace(3) %gep, float 4.0 seq_cst
+ ret float %result
+}
+
+define void @local_atomic_fmin_noret_f32(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmin_noret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v1, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, 4.0, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB2_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_noret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX940-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB2_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_noret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB2_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_noret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX10-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB2_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_noret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX90A-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB2_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_noret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX908-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB2_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_noret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX8-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB2_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_noret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX7-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_noret_f32:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX6-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB2_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(3) %ptr, float 4.0 seq_cst
+ ret void
+}
+
+define void @local_atomic_fmin_noret_f32__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmin_noret_f32__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65532
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v1, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, 4.0, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB3_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_noret_f32__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX940-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB3_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_noret_f32__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65532
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB3_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_noret_f32__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX10-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB3_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_noret_f32__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX90A-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB3_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_noret_f32__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX908-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB3_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_noret_f32__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX8-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB3_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_noret_f32__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX7-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB3_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_noret_f32__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX6-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB3_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(3) %ptr, i32 16383
+ %unused = atomicrmw fmin ptr addrspace(3) %gep, float 4.0 seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; double
+; --------------------------------------------------------------------
+
+define double @local_atomic_fmin_ret_f64(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmin_ret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v0
+; GFX12-NEXT: ds_load_b64 v[0:1], v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[0:1], v[3:4], v[3:4]
+; GFX12-NEXT: v_min_num_f64_e32 v[0:1], 4.0, v[0:1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4]
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB4_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_ret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v0
+; GFX940-NEXT: ds_read_b64 v[0:1], v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[0:1]
+; GFX940-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
+; GFX940-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX940-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[4:5], v[0:1]
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB4_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_ret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v0
+; GFX11-NEXT: ds_load_b64 v[0:1], v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX11-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4]
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB4_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_ret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v0
+; GFX10-NEXT: ds_read_b64 v[0:1], v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v1
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX10-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB4_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_ret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v0
+; GFX90A-NEXT: ds_read_b64 v[0:1], v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
+; GFX90A-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX90A-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[4:5], v[0:1]
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB4_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_ret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v0
+; GFX908-NEXT: ds_read_b64 v[0:1], v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v1
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX908-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX908-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB4_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_ret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: v_mov_b32_e32 v2, v0
+; GFX8-NEXT: ds_read_b64 v[0:1], v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v1
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX8-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX8-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB4_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_ret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: v_mov_b32_e32 v2, v0
+; GFX7-NEXT: ds_read_b64 v[0:1], v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX7-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX7-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB4_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_ret_f64:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: v_mov_b32_e32 v2, v0
+; GFX6-NEXT: ds_read_b64 v[0:1], v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX6-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX6-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB4_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(3) %ptr, double 4.0 seq_cst
+ ret double %result
+}
+
+define double @local_atomic_fmin_ret_f64__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmin_ret_f64__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v0
+; GFX12-NEXT: ds_load_b64 v[0:1], v0 offset:65528
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[0:1], v[3:4], v[3:4]
+; GFX12-NEXT: v_min_num_f64_e32 v[0:1], 4.0, v[0:1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4] offset:65528
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB5_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_ret_f64__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v0
+; GFX940-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[0:1]
+; GFX940-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
+; GFX940-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX940-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[4:5], v[0:1] offset:65528
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB5_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_ret_f64__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v0
+; GFX11-NEXT: ds_load_b64 v[0:1], v0 offset:65528
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX11-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4] offset:65528
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB5_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_ret_f64__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v0
+; GFX10-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v1
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX10-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] offset:65528
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB5_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_ret_f64__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v0
+; GFX90A-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
+; GFX90A-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX90A-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[4:5], v[0:1] offset:65528
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB5_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_ret_f64__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v0
+; GFX908-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v1
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX908-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX908-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] offset:65528
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_ret_f64__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: v_mov_b32_e32 v2, v0
+; GFX8-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v1
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX8-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX8-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] offset:65528
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_ret_f64__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: v_mov_b32_e32 v2, v0
+; GFX7-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX7-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX7-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] offset:65528
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_ret_f64__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v2, vcc, 0xfff8, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b64 v[0:1], v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: v_max_f64 v[0:1], v[3:4], v[3:4]
+; GFX6-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX6-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB5_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(3) %ptr, i32 8191
+ %result = atomicrmw fmin ptr addrspace(3) %gep, double 4.0 seq_cst
+ ret double %result
+}
+
+define void @local_atomic_fmin_noret_f64(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmin_noret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b64 v[1:2], v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_max_num_f64_e32 v[3:4], v[1:2], v[1:2]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f64_e32 v[3:4], 4.0, v[3:4]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2]
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX12-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB6_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_noret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b64 v[2:3], v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX940-NEXT: v_min_f64 v[4:5], v[4:5], 4.0
+; GFX940-NEXT: ds_cmpst_rtn_b64 v[4:5], v0, v[2:3], v[4:5]
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], v[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB6_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_noret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b64 v[1:2], v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2]
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX11-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB6_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_noret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b64 v[1:2], v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX10-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX10-NEXT: v_mov_b32_e32 v1, v3
+; GFX10-NEXT: v_mov_b32_e32 v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB6_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_noret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b64 v[2:3], v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX90A-NEXT: v_min_f64 v[4:5], v[4:5], 4.0
+; GFX90A-NEXT: ds_cmpst_rtn_b64 v[4:5], v0, v[2:3], v[4:5]
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB6_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_noret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b64 v[1:2], v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX908-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
+; GFX908-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX908-NEXT: v_mov_b32_e32 v1, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v2, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB6_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_noret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b64 v[1:2], v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX8-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
+; GFX8-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX8-NEXT: v_mov_b32_e32 v1, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v2, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB6_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_noret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b64 v[1:2], v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX7-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
+; GFX7-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v2, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB6_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_noret_f64:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b64 v[1:2], v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX6-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
+; GFX6-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX6-NEXT: v_mov_b32_e32 v1, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v2, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB6_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(3) %ptr, double 4.0 seq_cst
+ ret void
+}
+
+define void @local_atomic_fmin_noret_f64__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmin_noret_f64__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b64 v[1:2], v0 offset:65528
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_max_num_f64_e32 v[3:4], v[1:2], v[1:2]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f64_e32 v[3:4], 4.0, v[3:4]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2] offset:65528
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX12-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB7_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_noret_f64__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b64 v[2:3], v0 offset:65528
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX940-NEXT: v_min_f64 v[4:5], v[4:5], 4.0
+; GFX940-NEXT: ds_cmpst_rtn_b64 v[4:5], v0, v[2:3], v[4:5] offset:65528
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], v[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB7_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_noret_f64__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b64 v[1:2], v0 offset:65528
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2] offset:65528
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX11-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB7_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_noret_f64__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b64 v[1:2], v0 offset:65528
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX10-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX10-NEXT: v_mov_b32_e32 v1, v3
+; GFX10-NEXT: v_mov_b32_e32 v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB7_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_noret_f64__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b64 v[2:3], v0 offset:65528
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX90A-NEXT: v_min_f64 v[4:5], v[4:5], 4.0
+; GFX90A-NEXT: ds_cmpst_rtn_b64 v[4:5], v0, v[2:3], v[4:5] offset:65528
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB7_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_noret_f64__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b64 v[1:2], v0 offset:65528
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX908-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
+; GFX908-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX908-NEXT: v_mov_b32_e32 v1, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v2, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB7_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_noret_f64__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b64 v[1:2], v0 offset:65528
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX8-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
+; GFX8-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX8-NEXT: v_mov_b32_e32 v1, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v2, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB7_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_noret_f64__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b64 v[1:2], v0 offset:65528
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
+; GFX7-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
+; GFX7-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v2, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB7_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_noret_f64__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v2, vcc, 0xfff8, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b64 v[0:1], v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_max_f64 v[3:4], v[0:1], v[0:1]
+; GFX6-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
+; GFX6-NEXT: ds_cmpst_rtn_b64 v[3:4], v2, v[0:1], v[3:4]
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[0:1]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB7_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(3) %ptr, i32 8191
+ %unused = atomicrmw fmin ptr addrspace(3) %gep, double 4.0 seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; half
+; --------------------------------------------------------------------
+
+define half @local_atomic_fmin_ret_f16(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmin_ret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: ds_load_b32 v3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff
+; GFX12-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX12-NEXT: v_max_num_f16_e32 v3, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f16_e32 v3, 4.0, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB8_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_ret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v2, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v3
+; GFX940-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX940-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX940-NEXT: v_min_f16_e32 v3, 4.0, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX940-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB8_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_ret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: ds_load_b32 v3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff
+; GFX11-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX11-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f16_e32 v3, 4.0, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB8_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_ret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: ds_read_b32 v2, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX10-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_not_b32_e32 v3, v3
+; GFX10-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, v0, v4
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_min_f16_e32 v2, 4.0, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v1, v4, v2
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB8_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_ret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v3
+; GFX90A-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX90A-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX90A-NEXT: v_min_f16_e32 v3, 4.0, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX90A-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB8_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_ret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX908-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX908-NEXT: v_min_f16_e32 v3, 4.0, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_ret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX8-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX8-NEXT: v_min_f16_e32 v3, 4.0, v3
+; GFX8-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_ret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX7-NEXT: v_min_f32_e32 v3, 4.0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB8_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_ret_f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX6-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX6-NEXT: v_min_f32_e32 v3, 4.0, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB8_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(3) %ptr, half 4.0 seq_cst
+ ret half %result
+}
+
+define half @local_atomic_fmin_ret_f16__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmin_ret_f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX12-NEXT: ds_load_b32 v3, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX12-NEXT: v_max_num_f16_e32 v3, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f16_e32 v3, 4.0, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB9_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_ret_f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v3, v1
+; GFX940-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v0, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v3
+; GFX940-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX940-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX940-NEXT: v_min_f16_e32 v3, 4.0, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX940-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB9_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_ret_f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX11-NEXT: ds_load_b32 v3, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX11-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f16_e32 v3, 4.0, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB9_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_ret_f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX10-NEXT: ds_read_b32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v2, v2
+; GFX10-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX10-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX10-NEXT: v_min_f16_e32 v3, 4.0, v3
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB9_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_ret_f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v3, v1
+; GFX90A-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v0, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v3
+; GFX90A-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX90A-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX90A-NEXT: v_min_f16_e32 v3, 4.0, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX90A-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB9_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_ret_f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v3, v1
+; GFX908-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v0, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX908-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX908-NEXT: v_min_f16_e32 v3, 4.0, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB9_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_ret_f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffe, v0
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v0, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX8-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX8-NEXT: v_min_f16_e32 v3, 4.0, v3
+; GFX8-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB9_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_ret_f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX7-NEXT: v_min_f32_e32 v3, 4.0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB9_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_ret_f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX6-NEXT: v_min_f32_e32 v3, 4.0, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB9_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(3) %ptr, i32 32767
+ %result = atomicrmw fmin ptr addrspace(3) %gep, half 4.0 seq_cst
+ ret half %result
+}
+
+define void @local_atomic_fmin_noret_f16(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmin_noret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: ds_load_b32 v2, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX12-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_not_b32_e32 v3, v3
+; GFX12-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX12-NEXT: v_max_num_f16_e32 v4, v4, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f16_e32 v4, 4.0, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX12-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX12-NEXT: v_mov_b32_e32 v2, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB10_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_noret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v2, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX940-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX940-NEXT: v_min_f16_e32 v4, 4.0, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX940-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB10_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_noret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: ds_load_b32 v2, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX11-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_not_b32_e32 v3, v3
+; GFX11-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX11-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f16_e32 v4, 4.0, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX11-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX11-NEXT: v_mov_b32_e32 v2, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB10_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_noret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: ds_read_b32 v2, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX10-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_not_b32_e32 v3, v3
+; GFX10-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX10-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX10-NEXT: v_min_f16_e32 v4, 4.0, v4
+; GFX10-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v1, v2, v4
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX10-NEXT: v_mov_b32_e32 v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB10_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_noret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX90A-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX90A-NEXT: v_min_f16_e32 v4, 4.0, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB10_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_noret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX908-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX908-NEXT: v_min_f16_e32 v4, 4.0, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB10_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_noret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX8-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX8-NEXT: v_min_f16_e32 v4, 4.0, v4
+; GFX8-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB10_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_noret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX7-NEXT: v_min_f32_e32 v4, 4.0, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB10_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_noret_f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX6-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX6-NEXT: v_min_f32_e32 v4, 4.0, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX6-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB10_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(3) %ptr, half 4.0 seq_cst
+ ret void
+}
+
+define void @local_atomic_fmin_noret_f16__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmin_noret_f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX12-NEXT: ds_load_b32 v3, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v4, v4, v4
+; GFX12-NEXT: v_min_num_f16_e32 v4, 4.0, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB11_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_noret_f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX940-NEXT: ds_read_b32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v1, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX940-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX940-NEXT: v_min_f16_e32 v4, 4.0, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX940-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB11_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_noret_f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX11-NEXT: ds_load_b32 v3, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX11-NEXT: v_min_f16_e32 v4, 4.0, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB11_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_noret_f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX10-NEXT: ds_read_b32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v2, v2
+; GFX10-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX10-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX10-NEXT: v_min_f16_e32 v4, 4.0, v4
+; GFX10-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB11_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_noret_f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX90A-NEXT: ds_read_b32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX90A-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX90A-NEXT: v_min_f16_e32 v4, 4.0, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_noret_f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX908-NEXT: ds_read_b32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX908-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX908-NEXT: v_min_f16_e32 v4, 4.0, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_noret_f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0xfffe, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX8-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX8-NEXT: v_min_f16_e32 v4, 4.0, v4
+; GFX8-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_noret_f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX7-NEXT: v_min_f32_e32 v4, 4.0, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB11_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_noret_f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX6-NEXT: v_min_f32_e32 v4, 4.0, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX6-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB11_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(3) %ptr, i32 32767
+ %unused = atomicrmw fmin ptr addrspace(3) %gep, half 4.0 seq_cst
+ ret void
+}
+
+define half @local_atomic_fmin_ret_f16__offset__align4(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmin_ret_f16__offset__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v1, v2, v2
+; GFX12-NEXT: v_min_num_f16_e32 v1, 4.0, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX12-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB12_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_ret_f16__offset__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v1
+; GFX940-NEXT: v_max_f16_e32 v1, v2, v2
+; GFX940-NEXT: v_min_f16_e32 v1, 4.0, v1
+; GFX940-NEXT: v_and_or_b32 v1, v2, s2, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB12_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_ret_f16__offset__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v1, v2, v2
+; GFX11-NEXT: v_min_f16_e32 v1, 4.0, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB12_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_ret_f16__offset__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v1
+; GFX10-NEXT: v_max_f16_e32 v1, v2, v2
+; GFX10-NEXT: v_min_f16_e32 v1, 4.0, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX10-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB12_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v1
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_ret_f16__offset__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v1
+; GFX90A-NEXT: v_max_f16_e32 v1, v2, v2
+; GFX90A-NEXT: v_min_f16_e32 v1, 4.0, v1
+; GFX90A-NEXT: v_and_or_b32 v1, v2, s6, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB12_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_ret_f16__offset__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v1
+; GFX908-NEXT: v_max_f16_e32 v1, v2, v2
+; GFX908-NEXT: v_min_f16_e32 v1, 4.0, v1
+; GFX908-NEXT: v_and_or_b32 v1, v2, s6, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB12_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_ret_f16__offset__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v1
+; GFX8-NEXT: v_max_f16_e32 v1, v2, v2
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX8-NEXT: v_min_f16_e32 v1, 4.0, v1
+; GFX8-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB12_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_ret_f16__offset__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX7-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB12_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_ret_f16__offset__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffe, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX6-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB12_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(3) %ptr, i32 32767
+ %result = atomicrmw fmin ptr addrspace(3) %gep, half 4.0 seq_cst, align 4
+ ret half %result
+}
+
+define void @local_atomic_fmin_noret_f16__offset__align4(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmin_noret_f16__offset__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_max_num_f16_e32 v2, v1, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f16_e32 v2, 4.0, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB13_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_noret_f16__offset__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_max_f16_e32 v2, v1, v1
+; GFX940-NEXT: v_min_f16_e32 v2, 4.0, v2
+; GFX940-NEXT: v_and_or_b32 v2, v1, s2, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB13_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_noret_f16__offset__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_max_f16_e32 v2, v1, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f16_e32 v2, 4.0, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB13_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_noret_f16__offset__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_max_f16_e32 v2, v1, v1
+; GFX10-NEXT: v_min_f16_e32 v2, 4.0, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB13_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_noret_f16__offset__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_max_f16_e32 v2, v1, v1
+; GFX90A-NEXT: v_min_f16_e32 v2, 4.0, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v1, s6, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB13_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_noret_f16__offset__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_max_f16_e32 v2, v1, v1
+; GFX908-NEXT: v_min_f16_e32 v2, 4.0, v2
+; GFX908-NEXT: v_and_or_b32 v2, v1, s6, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB13_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_noret_f16__offset__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_max_f16_e32 v2, v1, v1
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX8-NEXT: v_min_f16_e32 v2, 4.0, v2
+; GFX8-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB13_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_noret_f16__offset__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX7-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB13_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_noret_f16__offset__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffe, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v1
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX6-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB13_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(3) %ptr, i32 32767
+ %unused = atomicrmw fmin ptr addrspace(3) %gep, half 4.0 seq_cst, align 4
+ ret void
+}
+
+; --------------------------------------------------------------------
+; bfloat
+; --------------------------------------------------------------------
+
+define bfloat @local_atomic_fmin_ret_bf16(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmin_ret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: ds_load_b32 v3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff
+; GFX12-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v3, 4.0, v3
+; GFX12-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB14_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_ret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v2, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v3
+; GFX940-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_min_f32_e32 v3, 4.0, v3
+; GFX940-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX940-NEXT: v_add3_u32 v5, v5, v3, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB14_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_ret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: ds_load_b32 v3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff
+; GFX11-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v3, 4.0, v3
+; GFX11-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB14_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_ret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: ds_read_b32 v3, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff
+; GFX10-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_not_b32_e32 v2, v2
+; GFX10-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_min_f32_e32 v3, 4.0, v3
+; GFX10-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX10-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB14_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_ret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v3
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_min_f32_e32 v3, 4.0, v3
+; GFX90A-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX90A-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_ret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_min_f32_e32 v3, 4.0, v3
+; GFX908-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_ret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f32_e32 v3, 4.0, v3
+; GFX8-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v3
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v3, v6, v7, vcc
+; GFX8-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX8-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_ret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v3, 4.0, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_ret_bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX6-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_min_f32_e32 v3, 4.0, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB14_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(3) %ptr, bfloat 4.0 seq_cst
+ ret bfloat %result
+}
+
+define bfloat @local_atomic_fmin_ret_bf16__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmin_ret_bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX12-NEXT: ds_load_b32 v3, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v3, 4.0, v3
+; GFX12-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB15_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_ret_bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v3, v1
+; GFX940-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v0, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v3
+; GFX940-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_min_f32_e32 v3, 4.0, v3
+; GFX940-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX940-NEXT: v_add3_u32 v5, v5, v3, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB15_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_ret_bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX11-NEXT: ds_load_b32 v3, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v3, 4.0, v3
+; GFX11-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB15_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_ret_bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX10-NEXT: ds_read_b32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v2, v2
+; GFX10-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_lshrrev_b32_sdwa v3, v1, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_min_f32_e32 v3, 4.0, v3
+; GFX10-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX10-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB15_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_ret_bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v3, v1
+; GFX90A-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v0, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v3
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_min_f32_e32 v3, 4.0, v3
+; GFX90A-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX90A-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB15_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_ret_bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v3, v1
+; GFX908-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v0, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_min_f32_e32 v3, 4.0, v3
+; GFX908-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB15_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_ret_bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffe, v0
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v0, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f32_e32 v3, 4.0, v3
+; GFX8-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v3
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v3, v6, v7, vcc
+; GFX8-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX8-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB15_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_ret_bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v3, 4.0, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB15_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_ret_bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_min_f32_e32 v3, 4.0, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB15_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(3) %ptr, i32 32767
+ %result = atomicrmw fmin ptr addrspace(3) %gep, bfloat 4.0 seq_cst
+ ret bfloat %result
+}
+
+define void @local_atomic_fmin_noret_bf16(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmin_noret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: ds_load_b32 v2, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX12-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_not_b32_e32 v3, v3
+; GFX12-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v4, 4.0, v4
+; GFX12-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX12-NEXT: v_mov_b32_e32 v2, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB16_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_noret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v2, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_min_f32_e32 v4, 4.0, v4
+; GFX940-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX940-NEXT: v_add3_u32 v5, v5, v4, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB16_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_noret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: ds_load_b32 v2, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX11-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_not_b32_e32 v3, v3
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v4, 4.0, v4
+; GFX11-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX11-NEXT: v_mov_b32_e32 v2, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB16_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_noret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: ds_read_b32 v2, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX10-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_not_b32_e32 v3, v3
+; GFX10-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v4, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_min_f32_e32 v4, 4.0, v4
+; GFX10-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX10-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v1, v2, v4
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX10-NEXT: v_mov_b32_e32 v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB16_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_noret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_min_f32_e32 v4, 4.0, v4
+; GFX90A-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX90A-NEXT: v_add3_u32 v5, v5, v4, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB16_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_noret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_min_f32_e32 v4, 4.0, v4
+; GFX908-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX908-NEXT: v_add3_u32 v5, v5, v4, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX908-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB16_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_noret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f32_e32 v4, 4.0, v4
+; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v4, v6, v7, vcc
+; GFX8-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX8-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB16_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_noret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_min_f32_e32 v4, 4.0, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB16_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_noret_bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX6-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_min_f32_e32 v4, 4.0, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX6-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB16_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(3) %ptr, bfloat 4.0 seq_cst
+ ret void
+}
+
+define void @local_atomic_fmin_noret_bf16__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmin_noret_bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX12-NEXT: ds_load_b32 v3, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: v_min_num_f32_e32 v4, 4.0, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX12-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX12-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB17_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_noret_bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX940-NEXT: ds_read_b32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v1, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_min_f32_e32 v4, 4.0, v4
+; GFX940-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX940-NEXT: v_add3_u32 v5, v5, v4, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB17_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_noret_bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX11-NEXT: ds_load_b32 v3, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: v_min_f32_e32 v4, 4.0, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX11-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB17_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_noret_bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX10-NEXT: ds_read_b32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v2, v2
+; GFX10-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_min_f32_e32 v4, 4.0, v4
+; GFX10-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX10-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB17_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_noret_bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX90A-NEXT: ds_read_b32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_min_f32_e32 v4, 4.0, v4
+; GFX90A-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX90A-NEXT: v_add3_u32 v5, v5, v4, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB17_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_noret_bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX908-NEXT: ds_read_b32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_min_f32_e32 v4, 4.0, v4
+; GFX908-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX908-NEXT: v_add3_u32 v5, v5, v4, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX908-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_noret_bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0xfffe, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f32_e32 v4, 4.0, v4
+; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v4, v6, v7, vcc
+; GFX8-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX8-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_noret_bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_min_f32_e32 v4, 4.0, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_noret_bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_min_f32_e32 v4, 4.0, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX6-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB17_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(3) %ptr, i32 32767
+ %unused = atomicrmw fmin ptr addrspace(3) %gep, bfloat 4.0 seq_cst
+ ret void
+}
+
+define bfloat @local_atomic_fmin_ret_bf16__offset__align4(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmin_ret_bf16__offset__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX12-NEXT: v_min_num_f32_e32 v1, 4.0, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX12-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB18_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_ret_bf16__offset__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX940-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX940-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX940-NEXT: v_add3_u32 v3, v3, v1, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX940-NEXT: v_and_or_b32 v1, v2, s3, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB18_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_ret_bf16__offset__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX11-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB18_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_ret_bf16__offset__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v1
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX10-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX10-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX10-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX10-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB18_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v1
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_ret_bf16__offset__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX90A-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX90A-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX90A-NEXT: v_add3_u32 v3, v3, v1, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX90A-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX90A-NEXT: v_and_or_b32 v1, v2, s7, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB18_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_ret_bf16__offset__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX908-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX908-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX908-NEXT: v_add3_u32 v3, v3, v1, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX908-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX908-NEXT: v_and_or_b32 v1, v2, s7, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB18_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_ret_bf16__offset__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX8-NEXT: v_bfe_u32 v4, v1, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v1
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
+; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v1
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v1, v4, v5, vcc
+; GFX8-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB18_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_ret_bf16__offset__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB18_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_ret_bf16__offset__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffe, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB18_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(3) %ptr, i32 32767
+ %result = atomicrmw fmin ptr addrspace(3) %gep, bfloat 4.0 seq_cst, align 4
+ ret bfloat %result
+}
+
+define void @local_atomic_fmin_noret_bf16__offset__align4(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fmin_noret_bf16__offset__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, 4.0, v2
+; GFX12-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB19_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_noret_bf16__offset__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX940-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX940-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v3, v3, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: v_and_or_b32 v2, v1, s3, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB19_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_noret_bf16__offset__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX11-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB19_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_noret_bf16__offset__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX10-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX10-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB19_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_noret_bf16__offset__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX90A-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX90A-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v3, v3, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v1, s7, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB19_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_noret_bf16__offset__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX908-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX908-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v3, v3, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: v_and_or_b32 v2, v1, s7, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB19_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_noret_bf16__offset__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX8-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX8-NEXT: v_bfe_u32 v4, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v2
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
+; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc
+; GFX8-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB19_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_noret_bf16__offset__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB19_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_noret_bf16__offset__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffe, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB19_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(3) %ptr, i32 32767
+ %unused = atomicrmw fmin ptr addrspace(3) %gep, bfloat 4.0 seq_cst, align 4
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x half>
+; --------------------------------------------------------------------
+
+define <2 x half> @local_atomic_fmin_ret_v2f16(ptr addrspace(3) %ptr, <2 x half> %val) {
+; GFX12-LABEL: local_atomic_fmin_ret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0
+; GFX12-NEXT: v_pk_max_num_f16 v1, v1, v1
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v2, v3, v3
+; GFX12-NEXT: v_pk_min_num_f16 v2, v2, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB20_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_ret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX940-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_min_f16 v2, v2, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB20_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_ret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0
+; GFX11-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX11-NEXT: v_pk_min_f16 v2, v2, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB20_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_ret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0
+; GFX10-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX10-NEXT: v_pk_min_f16 v2, v2, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB20_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_ret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX90A-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX90A-NEXT: v_pk_min_f16 v2, v2, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB20_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_ret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX908-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX908-NEXT: v_pk_min_f16 v2, v2, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB20_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_ret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v2, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v3, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v1, v1, v1
+; GFX8-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: v_max_f16_sdwa v2, v4, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v4, v4
+; GFX8-NEXT: v_min_f16_sdwa v2, v2, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v1
+; GFX8-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB20_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_ret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v1
+; GFX7-NEXT: v_min_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v5
+; GFX7-NEXT: v_or_b32_e32 v1, v6, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB20_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_ret_v2f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_or_b32_e32 v7, v2, v1
+; GFX6-NEXT: v_min_f32_e32 v5, v5, v3
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v5
+; GFX6-NEXT: v_or_b32_e32 v1, v6, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB20_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v0, v2
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fmin ptr addrspace(3) %ptr, <2 x half> %val seq_cst
+ ret <2 x half> %result
+}
+
+define <2 x half> @local_atomic_fmin_ret_v2f16__offset(ptr addrspace(3) %ptr, <2 x half> %val) {
+; GFX12-LABEL: local_atomic_fmin_ret_v2f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX12-NEXT: v_pk_max_num_f16 v1, v1, v1
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v2, v3, v3
+; GFX12-NEXT: v_pk_min_num_f16 v2, v2, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB21_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_ret_v2f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX940-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_min_f16 v2, v2, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB21_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_ret_v2f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX11-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX11-NEXT: v_pk_min_f16 v2, v2, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB21_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_ret_v2f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX10-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX10-NEXT: v_pk_min_f16 v2, v2, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB21_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_ret_v2f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX90A-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX90A-NEXT: v_pk_min_f16 v2, v2, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB21_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_ret_v2f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX908-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX908-NEXT: v_pk_min_f16 v2, v2, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB21_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_ret_v2f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v3, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v1, v1, v1
+; GFX8-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: v_max_f16_sdwa v2, v4, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v4, v4
+; GFX8-NEXT: v_min_f16_sdwa v2, v2, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v1
+; GFX8-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB21_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_ret_v2f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v1
+; GFX7-NEXT: v_min_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v5
+; GFX7-NEXT: v_or_b32_e32 v1, v6, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v1 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB21_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_ret_v2f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v3, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v4, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX6-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v5
+; GFX6-NEXT: v_or_b32_e32 v0, v6, v0
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v3, v7, v0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB21_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fmin ptr addrspace(3) %gep, <2 x half> %val seq_cst
+ ret <2 x half> %result
+}
+
+define void @local_atomic_fmin_noret_v2f16(ptr addrspace(3) %ptr, <2 x half> %val) {
+; GFX12-LABEL: local_atomic_fmin_noret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0
+; GFX12-NEXT: v_pk_max_num_f16 v1, v1, v1
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v3, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_min_num_f16 v3, v3, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX12-NEXT: v_mov_b32_e32 v2, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB22_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_noret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX940-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v3, v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_min_f16 v3, v3, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB22_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_noret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0
+; GFX11-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v3, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_min_f16 v3, v3, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX11-NEXT: v_mov_b32_e32 v2, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB22_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_noret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0
+; GFX10-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v3, v2, v2
+; GFX10-NEXT: v_pk_min_f16 v3, v3, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX10-NEXT: v_mov_b32_e32 v2, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB22_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_noret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX90A-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v3, v2, v2
+; GFX90A-NEXT: v_pk_min_f16 v3, v3, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB22_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_noret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX908-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v3, v2, v2
+; GFX908-NEXT: v_pk_min_f16 v3, v3, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v2, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB22_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_noret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v2, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v1, v1, v1
+; GFX8-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v4, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v3, v3
+; GFX8-NEXT: v_min_f16_sdwa v4, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v1
+; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB22_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_noret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_or_b32_e32 v7, v3, v4
+; GFX7-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_or_b32_e32 v3, v6, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB22_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_noret_v2f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_or_b32_e32 v7, v3, v4
+; GFX6-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_or_b32_e32 v3, v6, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB22_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(3) %ptr, <2 x half> %val seq_cst
+ ret void
+}
+
+define void @local_atomic_fmin_noret_v2f16__offset(ptr addrspace(3) %ptr, <2 x half> %val) {
+; GFX12-LABEL: local_atomic_fmin_noret_v2f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX12-NEXT: v_pk_max_num_f16 v1, v1, v1
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v3, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_min_num_f16 v3, v3, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX12-NEXT: v_mov_b32_e32 v2, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB23_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_noret_v2f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX940-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v3, v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_min_f16 v3, v3, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB23_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_noret_v2f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX11-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v3, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_min_f16 v3, v3, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX11-NEXT: v_mov_b32_e32 v2, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB23_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_noret_v2f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX10-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v3, v2, v2
+; GFX10-NEXT: v_pk_min_f16 v3, v3, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX10-NEXT: v_mov_b32_e32 v2, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB23_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_noret_v2f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX90A-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v3, v2, v2
+; GFX90A-NEXT: v_pk_min_f16 v3, v3, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB23_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_noret_v2f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v1, v1, v1
+; GFX908-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v3, v2, v2
+; GFX908-NEXT: v_pk_min_f16 v3, v3, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v2, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB23_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_noret_v2f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v2, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v1, v1, v1
+; GFX8-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v4, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v3, v3
+; GFX8-NEXT: v_min_f16_sdwa v4, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v1
+; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB23_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_noret_v2f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_or_b32_e32 v7, v3, v4
+; GFX7-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_or_b32_e32 v3, v6, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v3 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB23_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_noret_v2f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v4, v0
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_or_b32_e32 v7, v3, v4
+; GFX6-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_or_b32_e32 v3, v6, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB23_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fmin ptr addrspace(3) %gep, <2 x half> %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x bfloat>
+; --------------------------------------------------------------------
+
+define <2 x bfloat> @local_atomic_fmin_ret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
+; GFX12-LABEL: local_atomic_fmin_ret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX12-NEXT: v_min_num_f32_e32 v5, v5, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v6, v8, s0
+; GFX12-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB24_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_ret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v3
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX940-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v2, s4
+; GFX940-NEXT: v_add3_u32 v8, v8, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v5, v2, s5
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB24_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_ret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX11-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v6, v8, s0
+; GFX11-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB24_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_ret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v3
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX10-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v6, v8, s4
+; GFX10-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB24_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_ret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v3
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX90A-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v2, s8
+; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v5, v2, s9
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB24_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_ret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v3
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX908-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v6, v6, v2, s8
+; GFX908-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v2, v5, v2, s9
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB24_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_ret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v2, v0
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX8-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v3
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX8-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v2, v5, v2, 16
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB24_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_ret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
+; GFX7-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_alignbit_b32 v3, v3, v6, 16
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v0, v1, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB24_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_ret_v2bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
+; GFX6-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_alignbit_b32 v1, v1, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_alignbit_b32 v3, v3, v6, 16
+; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v0, v1, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB24_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(3) %ptr, <2 x bfloat> %val seq_cst
+ ret <2 x bfloat> %result
+}
+
+define <2 x bfloat> @local_atomic_fmin_ret_v2bf16__offset(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
+; GFX12-LABEL: local_atomic_fmin_ret_v2bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX12-NEXT: v_min_num_f32_e32 v5, v5, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v6, v8, s0
+; GFX12-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB25_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_ret_v2bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v3
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX940-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v2, s4
+; GFX940-NEXT: v_add3_u32 v8, v8, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v5, v2, s5
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB25_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_ret_v2bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX11-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v6, v8, s0
+; GFX11-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB25_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_ret_v2bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v3
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX10-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v6, v8, s4
+; GFX10-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB25_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_ret_v2bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v3
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX90A-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v2, s8
+; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v5, v2, s9
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB25_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_ret_v2bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v3
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX908-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v6, v6, v2, s8
+; GFX908-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v2, v5, v2, s9
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB25_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_ret_v2bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX8-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v3
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX8-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v2, v5, v2, 16
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB25_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_ret_v2bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
+; GFX7-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_alignbit_b32 v3, v3, v6, 16
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v0, v1, v3 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB25_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_ret_v2bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v0, v4
+; GFX6-NEXT: v_mov_b32_e32 v3, v1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX6-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v0
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v3
+; GFX6-NEXT: v_alignbit_b32 v0, v1, v0, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v5
+; GFX6-NEXT: v_alignbit_b32 v1, v1, v6, 16
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v4, v0, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v5
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB25_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fmin ptr addrspace(3) %gep, <2 x bfloat> %val seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @local_atomic_fmin_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
+; GFX12-LABEL: local_atomic_fmin_noret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v3, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_min_num_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
+; GFX12-NEXT: v_min_num_f32_e32 v4, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v4, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v4
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB26_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_noret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v3, v0
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX940-NEXT: v_min_f32_e32 v4, v4, v2
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX940-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v4, s4
+; GFX940-NEXT: v_add3_u32 v8, v8, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v4, v4
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[0:1]
+; GFX940-NEXT: v_perm_b32 v4, v5, v4, s5
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB26_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_noret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v3, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_min_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
+; GFX11-NEXT: v_min_f32_e32 v4, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v4, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v4
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB26_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_noret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX10-NEXT: v_min_f32_e32 v4, v4, v2
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX10-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v4, v4
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v4, v6, v8, s4
+; GFX10-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v4
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB26_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_noret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v3, v0
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX90A-NEXT: v_min_f32_e32 v4, v4, v2
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX90A-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v4, s8
+; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v4, v5, v4, s9
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB26_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_noret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v3, v0
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX908-NEXT: v_min_f32_e32 v4, v4, v2
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX908-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v6, v6, v4, s8
+; GFX908-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX908-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v4, v5, v4, s9
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB26_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_noret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v0
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX8-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX8-NEXT: v_min_f32_e32 v4, v4, v2
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v4, v5, v4, 16
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB26_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_noret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX7-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v1
+; GFX7-NEXT: v_alignbit_b32 v3, v4, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB26_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_noret_v2bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX6-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v1
+; GFX6-NEXT: v_alignbit_b32 v3, v4, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v0, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB26_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(3) %ptr, <2 x bfloat> %val seq_cst
+ ret void
+}
+
+define void @local_atomic_fmin_noret_v2bf16__ofset(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
+; GFX12-LABEL: local_atomic_fmin_noret_v2bf16__ofset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v3, v0 offset:65532
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_min_num_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
+; GFX12-NEXT: v_min_num_f32_e32 v4, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v4, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v4
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB27_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_noret_v2bf16__ofset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX940-NEXT: v_min_f32_e32 v4, v4, v2
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX940-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v4, s4
+; GFX940-NEXT: v_add3_u32 v8, v8, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v4, v4
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[0:1]
+; GFX940-NEXT: v_perm_b32 v4, v5, v4, s5
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB27_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_noret_v2bf16__ofset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v3, v0 offset:65532
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_min_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
+; GFX11-NEXT: v_min_f32_e32 v4, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v4, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v4
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB27_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_noret_v2bf16__ofset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX10-NEXT: v_min_f32_e32 v4, v4, v2
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX10-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v4, v4
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v4, v6, v8, s4
+; GFX10-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v4
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB27_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_noret_v2bf16__ofset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX90A-NEXT: v_min_f32_e32 v4, v4, v2
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX90A-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v4, s8
+; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v4, v5, v4, s9
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB27_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_noret_v2bf16__ofset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX908-NEXT: v_min_f32_e32 v4, v4, v2
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX908-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v6, v6, v4, s8
+; GFX908-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX908-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v4, v5, v4, s9
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB27_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_noret_v2bf16__ofset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX8-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX8-NEXT: v_min_f32_e32 v4, v4, v2
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v1
+; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v4, v5, v4, 16
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB27_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_noret_v2bf16__ofset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX7-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v1
+; GFX7-NEXT: v_alignbit_b32 v3, v4, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v3, v4 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB27_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_noret_v2bf16__ofset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX6-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v1
+; GFX6-NEXT: v_alignbit_b32 v3, v4, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v0, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB27_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fmin ptr addrspace(3) %gep, <2 x bfloat> %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; misc
+; --------------------------------------------------------------------
+
+define float @local_atomic_fmin_ret_f32__amdgpu_ignore_denormal_mode(ptr addrspace(3) %ptr) {
+; GFX12-LABEL: local_atomic_fmin_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v1, v2, v2
+; GFX12-NEXT: v_min_num_f32_e32 v1, 4.0, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB28_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v1
+; GFX940-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX940-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB28_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX11-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB28_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v1
+; GFX10-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX10-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB28_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v1
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v1
+; GFX90A-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX90A-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB28_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v1
+; GFX908-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX908-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB28_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v1
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB28_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB28_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v1
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX6-NEXT: v_min_f32_e32 v1, 4.0, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB28_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v0, v1
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(3) %ptr, float 4.0 seq_cst, !amdgpu.ignore.denormal.mode !0
+ ret float %result
+}
+
+define void @local_atomic_fmin_noret_f32__amdgpu_ignore_denormal_mode(ptr addrspace(3) %ptr) {
+; GFX12-LABEL: local_atomic_fmin_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v1, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, 4.0, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB29_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fmin_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX940-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB29_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fmin_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB29_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fmin_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX10-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB29_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fmin_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX90A-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB29_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fmin_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX908-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB29_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fmin_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX8-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB29_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fmin_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX7-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB29_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fmin_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX6-NEXT: v_min_f32_e32 v2, 4.0, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB29_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(3) %ptr, float 4.0 seq_cst, !amdgpu.ignore.denormal.mode !0
+ ret void
+}
+
+!0 = !{}
diff --git a/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fsub.ll b/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fsub.ll
new file mode 100644
index 0000000000000..760bb5253f3dd
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fsub.ll
@@ -0,0 +1,8226 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 < %s | FileCheck -check-prefix=GFX12 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx940 < %s | FileCheck -check-prefix=GFX940 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx90a < %s | FileCheck -check-prefix=GFX90A %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx908 < %s | FileCheck -check-prefix=GFX908 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tonga < %s | FileCheck -check-prefix=GFX8 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=hawaii < %s | FileCheck -check-prefix=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tahiti < %s | FileCheck -check-prefix=GFX6 %s
+
+; --------------------------------------------------------------------
+; float
+; --------------------------------------------------------------------
+
+define float @local_atomic_fsub_ret_f32(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fsub_ret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB0_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_ret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v1
+; GFX940-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB0_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_ret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB0_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_ret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v1
+; GFX10-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB0_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v1
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_ret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v1
+; GFX90A-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB0_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_ret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v1
+; GFX908-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB0_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_ret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v1
+; GFX8-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB0_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_ret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v1
+; GFX7-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB0_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_ret_f32:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v1
+; GFX6-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB0_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v0, v1
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr addrspace(3) %ptr, float 4.0 seq_cst
+ ret float %result
+}
+
+define float @local_atomic_fsub_ret_f32__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fsub_ret_f32__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65532
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB1_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_ret_f32__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v1
+; GFX940-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB1_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_ret_f32__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65532
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB1_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_ret_f32__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v1
+; GFX10-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB1_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v1
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_ret_f32__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v1
+; GFX90A-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB1_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_ret_f32__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v1
+; GFX908-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB1_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_ret_f32__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v1
+; GFX8-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB1_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_ret_f32__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v1
+; GFX7-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB1_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_ret_f32__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v1, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v0, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v0
+; GFX6-NEXT: v_add_f32_e32 v0, -4.0, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v0, v1, v2, v0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v0, v2
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB1_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fsub ptr addrspace(3) %gep, float 4.0 seq_cst
+ ret float %result
+}
+
+define void @local_atomic_fsub_noret_f32(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fsub_noret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB2_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_noret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB2_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_noret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB2_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_noret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB2_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_noret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB2_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_noret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB2_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_noret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB2_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_noret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_noret_f32:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB2_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr addrspace(3) %ptr, float 4.0 seq_cst
+ ret void
+}
+
+define void @local_atomic_fsub_noret_f32__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fsub_noret_f32__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65532
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB3_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_noret_f32__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB3_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_noret_f32__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65532
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB3_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_noret_f32__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB3_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_noret_f32__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB3_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_noret_f32__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB3_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_noret_f32__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB3_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_noret_f32__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65532
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB3_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_noret_f32__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB3_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(3) %ptr, i32 16383
+ %unused = atomicrmw fsub ptr addrspace(3) %gep, float 4.0 seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; double
+; --------------------------------------------------------------------
+
+define double @local_atomic_fsub_ret_f64(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fsub_ret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v0
+; GFX12-NEXT: ds_load_b64 v[0:1], v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_add_f64_e32 v[0:1], -4.0, v[3:4]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4]
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB4_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_ret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v0
+; GFX940-NEXT: ds_read_b64 v[0:1], v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[0:1]
+; GFX940-NEXT: v_add_f64 v[0:1], v[4:5], -4.0
+; GFX940-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[4:5], v[0:1]
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB4_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_ret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v0
+; GFX11-NEXT: ds_load_b64 v[0:1], v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_f64 v[0:1], v[3:4], -4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4]
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB4_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_ret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v0
+; GFX10-NEXT: ds_read_b64 v[0:1], v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v1
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_add_f64 v[0:1], v[3:4], -4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB4_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_ret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v0
+; GFX90A-NEXT: ds_read_b64 v[0:1], v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: v_add_f64 v[0:1], v[4:5], -4.0
+; GFX90A-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[4:5], v[0:1]
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB4_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_ret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v0
+; GFX908-NEXT: ds_read_b64 v[0:1], v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v1
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_add_f64 v[0:1], v[3:4], -4.0
+; GFX908-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB4_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_ret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: v_mov_b32_e32 v2, v0
+; GFX8-NEXT: ds_read_b64 v[0:1], v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v1
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_add_f64 v[0:1], v[3:4], -4.0
+; GFX8-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB4_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_ret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: v_mov_b32_e32 v2, v0
+; GFX7-NEXT: ds_read_b64 v[0:1], v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_add_f64 v[0:1], v[3:4], -4.0
+; GFX7-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB4_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_ret_f64:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: v_mov_b32_e32 v2, v0
+; GFX6-NEXT: ds_read_b64 v[0:1], v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: v_add_f64 v[0:1], v[3:4], -4.0
+; GFX6-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB4_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr addrspace(3) %ptr, double 4.0 seq_cst
+ ret double %result
+}
+
+define double @local_atomic_fsub_ret_f64__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fsub_ret_f64__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v0
+; GFX12-NEXT: ds_load_b64 v[0:1], v0 offset:65528
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_add_f64_e32 v[0:1], -4.0, v[3:4]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4] offset:65528
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB5_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_ret_f64__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v0
+; GFX940-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[0:1]
+; GFX940-NEXT: v_add_f64 v[0:1], v[4:5], -4.0
+; GFX940-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[4:5], v[0:1] offset:65528
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB5_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_ret_f64__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v0
+; GFX11-NEXT: ds_load_b64 v[0:1], v0 offset:65528
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_f64 v[0:1], v[3:4], -4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4] offset:65528
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB5_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_ret_f64__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v0
+; GFX10-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v1
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_add_f64 v[0:1], v[3:4], -4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] offset:65528
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB5_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_ret_f64__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v0
+; GFX90A-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: v_add_f64 v[0:1], v[4:5], -4.0
+; GFX90A-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[4:5], v[0:1] offset:65528
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB5_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_ret_f64__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v0
+; GFX908-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v1
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_add_f64 v[0:1], v[3:4], -4.0
+; GFX908-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] offset:65528
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_ret_f64__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: v_mov_b32_e32 v2, v0
+; GFX8-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v1
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_add_f64 v[0:1], v[3:4], -4.0
+; GFX8-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] offset:65528
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_ret_f64__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: v_mov_b32_e32 v2, v0
+; GFX7-NEXT: ds_read_b64 v[0:1], v0 offset:65528
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_add_f64 v[0:1], v[3:4], -4.0
+; GFX7-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] offset:65528
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_ret_f64__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v2, vcc, 0xfff8, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b64 v[0:1], v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: v_add_f64 v[0:1], v[3:4], -4.0
+; GFX6-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB5_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(3) %ptr, i32 8191
+ %result = atomicrmw fsub ptr addrspace(3) %gep, double 4.0 seq_cst
+ ret double %result
+}
+
+define void @local_atomic_fsub_noret_f64(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fsub_noret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b64 v[1:2], v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_add_f64_e32 v[3:4], -4.0, v[1:2]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2]
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX12-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB6_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_noret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b64 v[2:3], v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_add_f64 v[4:5], v[2:3], -4.0
+; GFX940-NEXT: ds_cmpst_rtn_b64 v[4:5], v0, v[2:3], v[4:5]
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], v[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB6_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_noret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b64 v[1:2], v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_add_f64 v[3:4], v[1:2], -4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2]
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX11-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB6_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_noret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b64 v[1:2], v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_add_f64 v[3:4], v[1:2], -4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX10-NEXT: v_mov_b32_e32 v1, v3
+; GFX10-NEXT: v_mov_b32_e32 v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB6_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_noret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b64 v[2:3], v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_add_f64 v[4:5], v[2:3], -4.0
+; GFX90A-NEXT: ds_cmpst_rtn_b64 v[4:5], v0, v[2:3], v[4:5]
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB6_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_noret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b64 v[1:2], v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_add_f64 v[3:4], v[1:2], -4.0
+; GFX908-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX908-NEXT: v_mov_b32_e32 v1, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v2, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB6_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_noret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b64 v[1:2], v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_add_f64 v[3:4], v[1:2], -4.0
+; GFX8-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX8-NEXT: v_mov_b32_e32 v1, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v2, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB6_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_noret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b64 v[1:2], v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_add_f64 v[3:4], v[1:2], -4.0
+; GFX7-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v2, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB6_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_noret_f64:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b64 v[1:2], v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_add_f64 v[3:4], v[1:2], -4.0
+; GFX6-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX6-NEXT: v_mov_b32_e32 v1, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v2, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB6_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr addrspace(3) %ptr, double 4.0 seq_cst
+ ret void
+}
+
+define void @local_atomic_fsub_noret_f64__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fsub_noret_f64__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b64 v[1:2], v0 offset:65528
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_add_f64_e32 v[3:4], -4.0, v[1:2]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2] offset:65528
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX12-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB7_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_noret_f64__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b64 v[2:3], v0 offset:65528
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_add_f64 v[4:5], v[2:3], -4.0
+; GFX940-NEXT: ds_cmpst_rtn_b64 v[4:5], v0, v[2:3], v[4:5] offset:65528
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], v[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB7_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_noret_f64__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b64 v[1:2], v0 offset:65528
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_add_f64 v[3:4], v[1:2], -4.0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2] offset:65528
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX11-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB7_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_noret_f64__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b64 v[1:2], v0 offset:65528
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_add_f64 v[3:4], v[1:2], -4.0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
+; GFX10-NEXT: v_mov_b32_e32 v1, v3
+; GFX10-NEXT: v_mov_b32_e32 v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB7_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_noret_f64__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b64 v[2:3], v0 offset:65528
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_add_f64 v[4:5], v[2:3], -4.0
+; GFX90A-NEXT: ds_cmpst_rtn_b64 v[4:5], v0, v[2:3], v[4:5] offset:65528
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB7_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_noret_f64__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b64 v[1:2], v0 offset:65528
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_add_f64 v[3:4], v[1:2], -4.0
+; GFX908-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX908-NEXT: v_mov_b32_e32 v1, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v2, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB7_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_noret_f64__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b64 v[1:2], v0 offset:65528
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_add_f64 v[3:4], v[1:2], -4.0
+; GFX8-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX8-NEXT: v_mov_b32_e32 v1, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v2, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB7_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_noret_f64__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b64 v[1:2], v0 offset:65528
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_add_f64 v[3:4], v[1:2], -4.0
+; GFX7-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v2, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB7_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_noret_f64__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v2, vcc, 0xfff8, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b64 v[0:1], v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_add_f64 v[3:4], v[0:1], -4.0
+; GFX6-NEXT: ds_cmpst_rtn_b64 v[3:4], v2, v[0:1], v[3:4]
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[0:1]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB7_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(3) %ptr, i32 8191
+ %unused = atomicrmw fsub ptr addrspace(3) %gep, double 4.0 seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; half
+; --------------------------------------------------------------------
+
+define half @local_atomic_fsub_ret_f16(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fsub_ret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: ds_load_b32 v2, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX12-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_not_b32_e32 v3, v3
+; GFX12-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v0, v4
+; GFX12-NEXT: v_add_f16_e32 v2, -4.0, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v0, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v1, v2, v4
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB8_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v0, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_ret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v2, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_and_b32_e32 v0, 24, v3
+; GFX940-NEXT: v_lshlrev_b32_e64 v3, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v3, v3
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v2
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, v0, v4
+; GFX940-NEXT: v_add_f16_e32 v2, -4.0, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, v0, v2
+; GFX940-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v1, v4, v2
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB8_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v0, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_ret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: ds_load_b32 v2, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX11-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_not_b32_e32 v3, v3
+; GFX11-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v0, v4
+; GFX11-NEXT: v_add_f16_e32 v2, -4.0, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v0, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v1, v2, v4
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB8_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v0, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_ret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: ds_read_b32 v2, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX10-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_not_b32_e32 v3, v3
+; GFX10-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, v0, v4
+; GFX10-NEXT: v_add_f16_e32 v2, -4.0, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v1, v4, v2
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB8_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_ret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v2, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_and_b32_e32 v0, 24, v3
+; GFX90A-NEXT: v_lshlrev_b32_e64 v3, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v3, v3
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v2
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, v0, v4
+; GFX90A-NEXT: v_add_f16_e32 v2, -4.0, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, v0, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v1, v4, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB8_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_ret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v2, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_and_b32_e32 v0, 24, v3
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v3, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, v0, v4
+; GFX908-NEXT: v_add_f16_e32 v2, -4.0, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, v0, v2
+; GFX908-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v1, v4, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v0, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_ret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX8-NEXT: v_add_f16_e32 v3, -4.0, v3
+; GFX8-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_ret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX7-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB8_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_ret_f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX6-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX6-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB8_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr addrspace(3) %ptr, half 4.0 seq_cst
+ ret half %result
+}
+
+define half @local_atomic_fsub_ret_f16__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fsub_ret_f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX12-NEXT: ds_load_b32 v3, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX12-NEXT: v_add_f16_e32 v3, -4.0, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB9_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_ret_f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v2, v1
+; GFX940-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v3, v0, s0
+; GFX940-NEXT: v_not_b32_e32 v3, v3
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v2
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, v0, v4
+; GFX940-NEXT: v_add_f16_e32 v2, -4.0, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, v0, v2
+; GFX940-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v1, v4, v2
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB9_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v0, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_ret_f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX11-NEXT: ds_load_b32 v3, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX11-NEXT: v_add_f16_e32 v3, -4.0, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB9_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_ret_f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX10-NEXT: ds_read_b32 v2, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v1, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v3, v3
+; GFX10-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, v1, v4
+; GFX10-NEXT: v_add_f16_e32 v2, -4.0, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB9_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v1, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_ret_f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v2, v1
+; GFX90A-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v3, v0, s4
+; GFX90A-NEXT: v_not_b32_e32 v3, v3
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v2
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, v0, v4
+; GFX90A-NEXT: v_add_f16_e32 v2, -4.0, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, v0, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v1, v4, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB9_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_ret_f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v2, v1
+; GFX908-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v0, s4
+; GFX908-NEXT: v_not_b32_e32 v3, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, v0, v4
+; GFX908-NEXT: v_add_f16_e32 v2, -4.0, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, v0, v2
+; GFX908-NEXT: v_and_or_b32 v2, v4, v3, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v1, v4, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB9_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v0, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_ret_f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffe, v0
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v0, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX8-NEXT: v_add_f16_e32 v3, -4.0, v3
+; GFX8-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB9_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_ret_f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX7-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB9_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_ret_f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX6-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB9_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(3) %ptr, i32 32767
+ %result = atomicrmw fsub ptr addrspace(3) %gep, half 4.0 seq_cst
+ ret half %result
+}
+
+define void @local_atomic_fsub_noret_f16(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fsub_noret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: ds_load_b32 v2, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX12-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_not_b32_e32 v3, v3
+; GFX12-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX12-NEXT: v_add_f16_e32 v4, -4.0, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX12-NEXT: v_mov_b32_e32 v2, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB10_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_noret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v2, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX940-NEXT: v_add_f16_e32 v4, -4.0, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX940-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB10_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_noret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: ds_load_b32 v2, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX11-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_not_b32_e32 v3, v3
+; GFX11-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX11-NEXT: v_add_f16_e32 v4, -4.0, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX11-NEXT: v_mov_b32_e32 v2, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB10_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_noret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: ds_read_b32 v2, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX10-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_not_b32_e32 v3, v3
+; GFX10-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX10-NEXT: v_add_f16_e32 v4, -4.0, v4
+; GFX10-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v1, v2, v4
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX10-NEXT: v_mov_b32_e32 v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB10_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_noret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX90A-NEXT: v_add_f16_e32 v4, -4.0, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB10_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_noret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX908-NEXT: v_add_f16_e32 v4, -4.0, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB10_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_noret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX8-NEXT: v_add_f16_e32 v4, -4.0, v4
+; GFX8-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB10_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_noret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX7-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB10_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_noret_f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX6-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX6-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX6-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB10_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr addrspace(3) %ptr, half 4.0 seq_cst
+ ret void
+}
+
+define void @local_atomic_fsub_noret_f16__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fsub_noret_f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX12-NEXT: ds_load_b32 v3, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f16_e32 v4, -4.0, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX12-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB11_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_noret_f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX940-NEXT: ds_read_b32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v1, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX940-NEXT: v_add_f16_e32 v4, -4.0, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX940-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB11_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_noret_f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX11-NEXT: ds_load_b32 v3, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f16_e32 v4, -4.0, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX11-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB11_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_noret_f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX10-NEXT: ds_read_b32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v2, v2
+; GFX10-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX10-NEXT: v_add_f16_e32 v4, -4.0, v4
+; GFX10-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB11_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_noret_f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX90A-NEXT: ds_read_b32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX90A-NEXT: v_add_f16_e32 v4, -4.0, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_noret_f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX908-NEXT: ds_read_b32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX908-NEXT: v_add_f16_e32 v4, -4.0, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_noret_f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0xfffe, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX8-NEXT: v_add_f16_e32 v4, -4.0, v4
+; GFX8-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_noret_f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX7-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB11_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_noret_f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX6-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX6-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB11_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(3) %ptr, i32 32767
+ %unused = atomicrmw fsub ptr addrspace(3) %gep, half 4.0 seq_cst
+ ret void
+}
+
+define half @local_atomic_fsub_ret_f16__offset__align4(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fsub_ret_f16__offset__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f16_e32 v1, -4.0, v2
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB12_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_ret_f16__offset__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v1
+; GFX940-NEXT: v_add_f16_e32 v1, -4.0, v2
+; GFX940-NEXT: v_and_or_b32 v1, v2, s2, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB12_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_ret_f16__offset__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f16_e32 v1, -4.0, v2
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB12_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_ret_f16__offset__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v1
+; GFX10-NEXT: v_add_f16_e32 v1, -4.0, v2
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX10-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB12_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v1
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_ret_f16__offset__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v1
+; GFX90A-NEXT: v_add_f16_e32 v1, -4.0, v2
+; GFX90A-NEXT: v_and_or_b32 v1, v2, s6, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB12_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_ret_f16__offset__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v1
+; GFX908-NEXT: v_add_f16_e32 v1, -4.0, v2
+; GFX908-NEXT: v_and_or_b32 v1, v2, s6, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB12_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_ret_f16__offset__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v1
+; GFX8-NEXT: v_add_f16_e32 v1, -4.0, v2
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX8-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB12_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_ret_f16__offset__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX7-NEXT: v_add_f32_e32 v1, -4.0, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB12_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_ret_f16__offset__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffe, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX6-NEXT: v_add_f32_e32 v1, -4.0, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB12_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(3) %ptr, i32 32767
+ %result = atomicrmw fsub ptr addrspace(3) %gep, half 4.0 seq_cst, align 4
+ ret half %result
+}
+
+define void @local_atomic_fsub_noret_f16__offset__align4(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fsub_noret_f16__offset__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_add_f16_e32 v2, -4.0, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB13_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_noret_f16__offset__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_add_f16_e32 v2, -4.0, v1
+; GFX940-NEXT: v_and_or_b32 v2, v1, s2, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB13_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_noret_f16__offset__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_add_f16_e32 v2, -4.0, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB13_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_noret_f16__offset__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_add_f16_e32 v2, -4.0, v1
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB13_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_noret_f16__offset__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_add_f16_e32 v2, -4.0, v1
+; GFX90A-NEXT: v_and_or_b32 v2, v1, s6, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB13_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_noret_f16__offset__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_add_f16_e32 v2, -4.0, v1
+; GFX908-NEXT: v_and_or_b32 v2, v1, s6, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB13_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_noret_f16__offset__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_add_f16_e32 v2, -4.0, v1
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX8-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB13_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_noret_f16__offset__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX7-NEXT: v_add_f32_e32 v2, -4.0, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB13_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_noret_f16__offset__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffe, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v1
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX6-NEXT: v_add_f32_e32 v2, -4.0, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB13_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(3) %ptr, i32 32767
+ %unused = atomicrmw fsub ptr addrspace(3) %gep, half 4.0 seq_cst, align 4
+ ret void
+}
+
+; --------------------------------------------------------------------
+; bfloat
+; --------------------------------------------------------------------
+
+define bfloat @local_atomic_fsub_ret_bf16(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fsub_ret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: ds_load_b32 v3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff
+; GFX12-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX12-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB14_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_ret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v2, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v3
+; GFX940-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX940-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX940-NEXT: v_add3_u32 v5, v5, v3, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB14_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_ret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: ds_load_b32 v3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff
+; GFX11-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX11-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB14_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_ret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: ds_read_b32 v3, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff
+; GFX10-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_not_b32_e32 v2, v2
+; GFX10-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX10-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX10-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB14_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_ret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v3
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX90A-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX90A-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_ret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX908-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_ret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX8-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v3
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v3, v6, v7, vcc
+; GFX8-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX8-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_ret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_ret_bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX6-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB14_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr addrspace(3) %ptr, bfloat 4.0 seq_cst
+ ret bfloat %result
+}
+
+define bfloat @local_atomic_fsub_ret_bf16__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fsub_ret_bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX12-NEXT: ds_load_b32 v3, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX12-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB15_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_ret_bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v3, v1
+; GFX940-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v0, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v3
+; GFX940-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX940-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX940-NEXT: v_add3_u32 v5, v5, v3, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB15_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_ret_bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX11-NEXT: ds_load_b32 v3, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX11-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB15_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_ret_bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX10-NEXT: ds_read_b32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v2, v2
+; GFX10-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_lshrrev_b32_sdwa v3, v1, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX10-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX10-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB15_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_ret_bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v3, v1
+; GFX90A-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v0, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v3
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX90A-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX90A-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB15_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_ret_bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v0, 0xfffe, v0
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v3, v1
+; GFX908-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v0, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX908-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v3, v4, v2, v3
+; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB15_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_ret_bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffe, v0
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_and_b32_e32 v0, 3, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v0, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX8-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v3
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v3, v6, v7, vcc
+; GFX8-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX8-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB15_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_ret_bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB15_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_ret_bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_add_f32_e32 v3, -4.0, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_and_b32_e32 v5, v4, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB15_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v1, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(3) %ptr, i32 32767
+ %result = atomicrmw fsub ptr addrspace(3) %gep, bfloat 4.0 seq_cst
+ ret bfloat %result
+}
+
+define void @local_atomic_fsub_noret_bf16(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fsub_noret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: ds_load_b32 v2, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX12-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_not_b32_e32 v3, v3
+; GFX12-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX12-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX12-NEXT: v_mov_b32_e32 v2, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB16_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_noret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX940-NEXT: ds_read_b32 v3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v2, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX940-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX940-NEXT: v_add3_u32 v5, v5, v4, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB16_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_noret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: ds_load_b32 v2, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX11-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_not_b32_e32 v3, v3
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX11-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX11-NEXT: v_mov_b32_e32 v2, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB16_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_noret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: ds_read_b32 v2, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
+; GFX10-NEXT: v_and_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_not_b32_e32 v3, v3
+; GFX10-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v4, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX10-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX10-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v4, v2, v3, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v1, v2, v4
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX10-NEXT: v_mov_b32_e32 v2, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB16_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_noret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX90A-NEXT: ds_read_b32 v3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX90A-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX90A-NEXT: v_add3_u32 v5, v5, v4, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB16_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_noret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX908-NEXT: ds_read_b32 v3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX908-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX908-NEXT: v_add3_u32 v5, v5, v4, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX908-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB16_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_noret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v2, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v4, v6, v7, vcc
+; GFX8-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX8-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB16_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_noret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB16_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_noret_bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, -4, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 3, v0
+; GFX6-NEXT: v_and_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, v0, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX6-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB16_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr addrspace(3) %ptr, bfloat 4.0 seq_cst
+ ret void
+}
+
+define void @local_atomic_fsub_noret_bf16__offset(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fsub_noret_bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX12-NEXT: ds_load_b32 v3, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX12-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v2, v2
+; GFX12-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX12-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX12-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB17_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_noret_bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX940-NEXT: ds_read_b32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v2, v1, s0
+; GFX940-NEXT: v_not_b32_e32 v2, v2
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX940-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX940-NEXT: v_add3_u32 v5, v5, v4, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB17_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_noret_bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX11-NEXT: ds_load_b32 v3, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX11-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v2, v2
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX11-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB17_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_noret_bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v1, 0xfffe, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX10-NEXT: ds_read_b32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX10-NEXT: v_lshlrev_b32_e64 v2, v1, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v2, v2
+; GFX10-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX10-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX10-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB17_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_noret_bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX90A-NEXT: ds_read_b32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX90A-NEXT: v_not_b32_e32 v2, v2
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX90A-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX90A-NEXT: v_add3_u32 v5, v5, v4, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB17_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_noret_bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v1, 0xfffe, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX908-NEXT: ds_read_b32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX908-NEXT: v_not_b32_e32 v2, v2
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX908-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX908-NEXT: v_add3_u32 v5, v5, v4, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX908-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_noret_bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0xfffe, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v2, v1, s4
+; GFX8-NEXT: v_not_b32_e32 v2, v2
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v4, v6, v7, vcc
+; GFX8-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX8-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_noret_bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX7-NEXT: v_not_b32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_noret_bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v1, vcc, 0xfffe, v0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v1
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 3, v1
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v1
+; GFX6-NEXT: v_not_b32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, v1, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_add_f32_e32 v4, -4.0, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_and_b32_e32 v5, v3, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX6-NEXT: v_or_b32_e32 v4, v5, v4
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB17_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(3) %ptr, i32 32767
+ %unused = atomicrmw fsub ptr addrspace(3) %gep, bfloat 4.0 seq_cst
+ ret void
+}
+
+define bfloat @local_atomic_fsub_ret_bf16__offset__align4(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fsub_ret_bf16__offset__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX12-NEXT: v_add_f32_e32 v1, -4.0, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX12-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB18_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_ret_bf16__offset__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX940-NEXT: v_add_f32_e32 v1, -4.0, v1
+; GFX940-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX940-NEXT: v_add3_u32 v3, v3, v1, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX940-NEXT: v_and_or_b32 v1, v2, s3, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB18_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_ret_bf16__offset__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX11-NEXT: v_add_f32_e32 v1, -4.0, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB18_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_ret_bf16__offset__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v1
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX10-NEXT: v_add_f32_e32 v1, -4.0, v1
+; GFX10-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX10-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX10-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB18_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v1
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_ret_bf16__offset__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX90A-NEXT: v_add_f32_e32 v1, -4.0, v1
+; GFX90A-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX90A-NEXT: v_add3_u32 v3, v3, v1, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX90A-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX90A-NEXT: v_and_or_b32 v1, v2, s7, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB18_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_ret_bf16__offset__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX908-NEXT: v_add_f32_e32 v1, -4.0, v1
+; GFX908-NEXT: v_bfe_u32 v3, v1, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX908-NEXT: v_add3_u32 v3, v3, v1, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX908-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX908-NEXT: v_and_or_b32 v1, v2, s7, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB18_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_ret_bf16__offset__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_add_f32_e32 v1, -4.0, v1
+; GFX8-NEXT: v_bfe_u32 v4, v1, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v1
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
+; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v1
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v1, v4, v5, vcc
+; GFX8-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB18_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_ret_bf16__offset__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX7-NEXT: v_add_f32_e32 v1, -4.0, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB18_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_ret_bf16__offset__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffe, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX6-NEXT: v_add_f32_e32 v1, -4.0, v1
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB18_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(3) %ptr, i32 32767
+ %result = atomicrmw fsub ptr addrspace(3) %gep, bfloat 4.0 seq_cst, align 4
+ ret bfloat %result
+}
+
+define void @local_atomic_fsub_noret_bf16__offset__align4(ptr addrspace(3) %ptr) nounwind {
+; GFX12-LABEL: local_atomic_fsub_noret_bf16__offset__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v2, -4.0, v2
+; GFX12-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB19_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_noret_bf16__offset__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX940-NEXT: v_add_f32_e32 v2, -4.0, v2
+; GFX940-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v3, v3, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: v_and_or_b32 v2, v1, s3, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB19_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_noret_bf16__offset__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0 offset:65534
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v2, -4.0, v2
+; GFX11-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB19_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_noret_bf16__offset__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX10-NEXT: v_add_f32_e32 v2, -4.0, v2
+; GFX10-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB19_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_noret_bf16__offset__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX90A-NEXT: v_add_f32_e32 v2, -4.0, v2
+; GFX90A-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v3, v3, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v1, s7, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB19_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_noret_bf16__offset__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX908-NEXT: v_add_f32_e32 v2, -4.0, v2
+; GFX908-NEXT: v_bfe_u32 v3, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v4, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v3, v3, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: v_and_or_b32 v2, v1, s7, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB19_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_noret_bf16__offset__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX8-NEXT: v_add_f32_e32 v2, -4.0, v2
+; GFX8-NEXT: v_bfe_u32 v4, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v2
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
+; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc
+; GFX8-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB19_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_noret_bf16__offset__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0 offset:65534
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX7-NEXT: v_add_f32_e32 v2, -4.0, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB19_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_noret_bf16__offset__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffe, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX6-NEXT: v_add_f32_e32 v2, -4.0, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v3, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB19_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(3) %ptr, i32 32767
+ %unused = atomicrmw fsub ptr addrspace(3) %gep, bfloat 4.0 seq_cst, align 4
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x half>
+; --------------------------------------------------------------------
+
+define <2 x half> @local_atomic_fsub_ret_v2f16(ptr addrspace(3) %ptr, <2 x half> %val) {
+; GFX12-LABEL: local_atomic_fsub_ret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB20_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_ret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB20_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_ret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB20_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_ret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB20_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_ret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB20_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_ret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB20_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_ret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v2, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: v_sub_f16_sdwa v2, v3, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_sub_f16_e32 v4, v3, v1
+; GFX8-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB20_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_ret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v1
+; GFX7-NEXT: v_sub_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v5
+; GFX7-NEXT: v_or_b32_e32 v1, v6, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB20_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_ret_v2f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_or_b32_e32 v7, v2, v1
+; GFX6-NEXT: v_sub_f32_e32 v5, v5, v3
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v5
+; GFX6-NEXT: v_or_b32_e32 v1, v6, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB20_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v0, v2
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fsub ptr addrspace(3) %ptr, <2 x half> %val seq_cst
+ ret <2 x half> %result
+}
+
+define <2 x half> @local_atomic_fsub_ret_v2f16__offset(ptr addrspace(3) %ptr, <2 x half> %val) {
+; GFX12-LABEL: local_atomic_fsub_ret_v2f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB21_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_ret_v2f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB21_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_ret_v2f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB21_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_ret_v2f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB21_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_ret_v2f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB21_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_ret_v2f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB21_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_ret_v2f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: v_sub_f16_sdwa v2, v3, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_sub_f16_e32 v4, v3, v1
+; GFX8-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB21_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_ret_v2f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v1
+; GFX7-NEXT: v_sub_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v5
+; GFX7-NEXT: v_or_b32_e32 v1, v6, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v1 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB21_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_ret_v2f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v3, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v4, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX6-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v5
+; GFX6-NEXT: v_or_b32_e32 v0, v6, v0
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v3, v7, v0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB21_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fsub ptr addrspace(3) %gep, <2 x half> %val seq_cst
+ ret <2 x half> %result
+}
+
+define void @local_atomic_fsub_noret_v2f16(ptr addrspace(3) %ptr, <2 x half> %val) {
+; GFX12-LABEL: local_atomic_fsub_noret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX12-NEXT: v_mov_b32_e32 v2, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB22_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_noret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB22_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_noret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX11-NEXT: v_mov_b32_e32 v2, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB22_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_noret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX10-NEXT: v_mov_b32_e32 v2, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB22_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_noret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB22_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_noret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v2, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB22_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_noret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v2, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_sub_f16_sdwa v3, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_sub_f16_e32 v4, v2, v1
+; GFX8-NEXT: v_or_b32_e32 v3, v4, v3
+; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v2, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB22_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_noret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_or_b32_e32 v7, v3, v4
+; GFX7-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_or_b32_e32 v3, v6, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB22_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_noret_v2f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_or_b32_e32 v7, v3, v4
+; GFX6-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_or_b32_e32 v3, v6, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB22_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr addrspace(3) %ptr, <2 x half> %val seq_cst
+ ret void
+}
+
+define void @local_atomic_fsub_noret_v2f16__offset(ptr addrspace(3) %ptr, <2 x half> %val) {
+; GFX12-LABEL: local_atomic_fsub_noret_v2f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX12-NEXT: v_mov_b32_e32 v2, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB23_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_noret_v2f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB23_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_noret_v2f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX11-NEXT: v_mov_b32_e32 v2, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB23_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_noret_v2f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX10-NEXT: v_mov_b32_e32 v2, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB23_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_noret_v2f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB23_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_noret_v2f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
+; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v2, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB23_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_noret_v2f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_sub_f16_sdwa v3, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_sub_f16_e32 v4, v2, v1
+; GFX8-NEXT: v_or_b32_e32 v3, v4, v3
+; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v2, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB23_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_noret_v2f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_or_b32_e32 v7, v3, v4
+; GFX7-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_or_b32_e32 v3, v6, v3
+; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v3 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB23_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_noret_v2f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v4, v0
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v1
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_or_b32_e32 v7, v3, v4
+; GFX6-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_or_b32_e32 v3, v6, v3
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB23_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fsub ptr addrspace(3) %gep, <2 x half> %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x bfloat>
+; --------------------------------------------------------------------
+
+define <2 x bfloat> @local_atomic_fsub_ret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
+; GFX12-LABEL: local_atomic_fsub_ret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX12-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
+; GFX12-NEXT: v_sub_f32_e32 v2, v2, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v6, v8, s0
+; GFX12-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB24_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_ret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX940-NEXT: v_sub_f32_e32 v2, v2, v3
+; GFX940-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX940-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v2, s4
+; GFX940-NEXT: v_add3_u32 v8, v8, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v5, v2, s5
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB24_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_ret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX11-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
+; GFX11-NEXT: v_sub_f32_e32 v2, v2, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v6, v8, s0
+; GFX11-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB24_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_ret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX10-NEXT: v_sub_f32_e32 v2, v2, v3
+; GFX10-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX10-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v6, v8, s4
+; GFX10-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB24_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_ret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX90A-NEXT: v_sub_f32_e32 v2, v2, v3
+; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX90A-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v2, s8
+; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v5, v2, s9
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB24_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_ret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX908-NEXT: v_sub_f32_e32 v2, v2, v3
+; GFX908-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX908-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v6, v6, v2, s8
+; GFX908-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v2, v5, v2, s9
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB24_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_ret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v2, v0
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX8-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX8-NEXT: v_sub_f32_e32 v2, v2, v3
+; GFX8-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX8-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v2, v5, v2, 16
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB24_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_ret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_sub_f32_e32 v5, v5, v4
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v3, v3, v5, 16
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v0, v1, v3
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB24_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_ret_v2bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v3, v0
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_sub_f32_e32 v5, v5, v4
+; GFX6-NEXT: v_alignbit_b32 v1, v1, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v3, v3, v5, 16
+; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v0, v1, v3
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB24_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr addrspace(3) %ptr, <2 x bfloat> %val seq_cst
+ ret <2 x bfloat> %result
+}
+
+define <2 x bfloat> @local_atomic_fsub_ret_v2bf16__offset(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
+; GFX12-LABEL: local_atomic_fsub_ret_v2bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX12-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
+; GFX12-NEXT: v_sub_f32_e32 v2, v2, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v6, v8, s0
+; GFX12-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB25_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_ret_v2bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX940-NEXT: v_sub_f32_e32 v2, v2, v3
+; GFX940-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX940-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v2, s4
+; GFX940-NEXT: v_add3_u32 v8, v8, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v5, v2, s5
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB25_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_ret_v2bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v2, v0 offset:65532
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX11-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
+; GFX11-NEXT: v_sub_f32_e32 v2, v2, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v6, v8, s0
+; GFX11-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB25_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_ret_v2bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX10-NEXT: v_sub_f32_e32 v2, v2, v3
+; GFX10-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX10-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v6, v8, s4
+; GFX10-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB25_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_ret_v2bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX90A-NEXT: v_sub_f32_e32 v2, v2, v3
+; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX90A-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v2, s8
+; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v5, v2, s9
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB25_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_ret_v2bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX908-NEXT: v_sub_f32_e32 v2, v2, v3
+; GFX908-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX908-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v6, v6, v2, s8
+; GFX908-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v2, v5, v2, s9
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB25_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_ret_v2bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v2, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX8-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v4
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX8-NEXT: v_sub_f32_e32 v2, v2, v3
+; GFX8-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX8-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v2, v5, v2, 16
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB25_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_ret_v2bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_sub_f32_e32 v5, v5, v4
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v3, v3, v5, 16
+; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v0, v1, v3 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB25_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_ret_v2bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v0, v4
+; GFX6-NEXT: v_mov_b32_e32 v3, v1
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v0
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_sub_f32_e32 v5, v5, v3
+; GFX6-NEXT: v_alignbit_b32 v0, v1, v0, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v1, v1, v5, 16
+; GFX6-NEXT: ds_cmpst_rtn_b32 v5, v4, v0, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v5
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB25_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fsub ptr addrspace(3) %gep, <2 x bfloat> %val seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @local_atomic_fsub_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
+; GFX12-LABEL: local_atomic_fsub_noret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v3, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_sub_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
+; GFX12-NEXT: v_sub_f32_e32 v4, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v4, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v4
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB26_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_noret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v3, v0
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX940-NEXT: v_sub_f32_e32 v4, v4, v2
+; GFX940-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX940-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v4, s4
+; GFX940-NEXT: v_add3_u32 v8, v8, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v4, v4
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[0:1]
+; GFX940-NEXT: v_perm_b32 v4, v5, v4, s5
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB26_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_noret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v3, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_sub_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
+; GFX11-NEXT: v_sub_f32_e32 v4, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v4, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v4
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB26_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_noret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX10-NEXT: v_sub_f32_e32 v4, v4, v2
+; GFX10-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX10-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v4, v4
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v4, v6, v8, s4
+; GFX10-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v4
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB26_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_noret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v3, v0
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX90A-NEXT: v_sub_f32_e32 v4, v4, v2
+; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX90A-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v4, s8
+; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v4, v5, v4, s9
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB26_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_noret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v3, v0
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX908-NEXT: v_sub_f32_e32 v4, v4, v2
+; GFX908-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX908-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v6, v6, v4, s8
+; GFX908-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX908-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v4, v5, v4, s9
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB26_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_noret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v0
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX8-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX8-NEXT: v_sub_f32_e32 v4, v4, v2
+; GFX8-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v4, v5, v4, 16
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB26_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_noret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v4, v0
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX7-NEXT: v_alignbit_b32 v3, v3, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v5, 16
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB26_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_noret_v2bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v4, v0
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX6-NEXT: v_alignbit_b32 v3, v3, v4, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v5, 16
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB26_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr addrspace(3) %ptr, <2 x bfloat> %val seq_cst
+ ret void
+}
+
+define void @local_atomic_fsub_noret_v2bf16__ofset(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
+; GFX12-LABEL: local_atomic_fsub_noret_v2bf16__ofset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v3, v0 offset:65532
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_sub_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
+; GFX12-NEXT: v_sub_f32_e32 v4, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v4, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 offset:65532
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v4
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB27_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_noret_v2bf16__ofset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX940-NEXT: v_sub_f32_e32 v4, v4, v2
+; GFX940-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX940-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v4, s4
+; GFX940-NEXT: v_add3_u32 v8, v8, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v4, v4
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[0:1]
+; GFX940-NEXT: v_perm_b32 v4, v5, v4, s5
+; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB27_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_noret_v2bf16__ofset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v3, v0 offset:65532
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_sub_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
+; GFX11-NEXT: v_sub_f32_e32 v4, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v4, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 offset:65532
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v4
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB27_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_noret_v2bf16__ofset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX10-NEXT: v_sub_f32_e32 v4, v4, v2
+; GFX10-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX10-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v4, v4
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v4, v6, v8, s4
+; GFX10-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v4
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB27_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_noret_v2bf16__ofset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX90A-NEXT: v_sub_f32_e32 v4, v4, v2
+; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX90A-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v4, s8
+; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v4, v5, v4, s9
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB27_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_noret_v2bf16__ofset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX908-NEXT: v_sub_f32_e32 v4, v4, v2
+; GFX908-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX908-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v6, v6, v4, s8
+; GFX908-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX908-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v4, v5, v4, s9
+; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB27_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_noret_v2bf16__ofset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v3, v0 offset:65532
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX8-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX8-NEXT: v_sub_f32_e32 v4, v4, v2
+; GFX8-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v4, v5, v4, 16
+; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB27_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_noret_v2bf16__ofset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v4, v0 offset:65532
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX7-NEXT: v_alignbit_b32 v3, v3, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v5, 16
+; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB27_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_noret_v2bf16__ofset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffc, v0
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v4, v0
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_sub_f32_e32 v5, v5, v1
+; GFX6-NEXT: v_alignbit_b32 v3, v3, v4, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v5, 16
+; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB27_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(3) %ptr, i32 16383
+ %result = atomicrmw fsub ptr addrspace(3) %gep, <2 x bfloat> %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; misc
+; --------------------------------------------------------------------
+
+define float @local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode(ptr addrspace(3) %ptr) {
+; GFX12-LABEL: local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB28_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v1
+; GFX940-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX940-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB28_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB28_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v1
+; GFX10-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB28_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v1
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v1
+; GFX90A-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB28_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v1
+; GFX908-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB28_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v1
+; GFX8-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB28_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v1
+; GFX7-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB28_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v1
+; GFX6-NEXT: v_add_f32_e32 v1, -4.0, v2
+; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB28_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v0, v1
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr addrspace(3) %ptr, float 4.0 seq_cst, !amdgpu.ignore.denormal.mode !0
+ ret float %result
+}
+
+define void @local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode(ptr addrspace(3) %ptr) {
+; GFX12-LABEL: local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: ds_load_b32 v1, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1
+; GFX12-NEXT: s_wait_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SE
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB29_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: ds_read_b32 v1, v0
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB29_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: ds_load_b32 v1, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB29_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: ds_read_b32 v1, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB29_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: ds_read_b32 v1, v0
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB29_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: ds_read_b32 v1, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB29_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_b32 v1, v0
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB29_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_b32 v1, v0
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB29_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 m0, -1
+; GFX6-NEXT: ds_read_b32 v1, v0
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v2, -4.0, v1
+; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB29_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr addrspace(3) %ptr, float 4.0 seq_cst, !amdgpu.ignore.denormal.mode !0
+ ret void
+}
+
+!0 = !{}
diff --git a/llvm/test/CodeGen/AMDGPU/local-atomics-fp.ll b/llvm/test/CodeGen/AMDGPU/local-atomics-fp.ll
deleted file mode 100644
index 4373b76070e32..0000000000000
--- a/llvm/test/CodeGen/AMDGPU/local-atomics-fp.ll
+++ /dev/null
@@ -1,3931 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
-; RUN: llc -mtriple=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=VI %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx908 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX9,GFX908 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx90a -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX9,GFX90A %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx940 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX9,GFX940 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX12 %s
-; RUN: llc -mtriple=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX7 %s
-; RUN: llc -mtriple=amdgcn -mcpu=hawaii -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX8 %s
-
-define float @lds_atomic_fadd_ret_f32(ptr addrspace(3) %ptr) nounwind {
-; VI-LABEL: lds_atomic_fadd_ret_f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v1, 4.0
-; VI-NEXT: s_mov_b32 m0, -1
-; VI-NEXT: ds_add_rtn_f32 v0, v0, v1
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: lds_atomic_fadd_ret_f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v1, 4.0
-; GFX9-NEXT: ds_add_rtn_f32 v0, v0, v1
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: lds_atomic_fadd_ret_f32:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v1, 4.0
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: ds_add_rtn_f32 v0, v0, v1
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SE
-; GFX12-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX7-LABEL: lds_atomic_fadd_ret_f32:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: s_mov_b32 m0, -1
-; GFX7-NEXT: ds_read_b32 v1, v0
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: .LBB0_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_mov_b32_e32 v2, v1
-; GFX7-NEXT: v_add_f32_e32 v1, 4.0, v2
-; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB0_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: v_mov_b32_e32 v0, v1
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: lds_atomic_fadd_ret_f32:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: s_mov_b32 m0, -1
-; GFX8-NEXT: ds_read_b32 v1, v0
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB0_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_mov_b32_e32 v2, v1
-; GFX8-NEXT: v_add_f32_e32 v1, 4.0, v2
-; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB0_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v0, v1
-; GFX8-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr addrspace(3) %ptr, float 4.0 seq_cst
- ret float %result
-}
-
-define void @lds_atomic_fadd_noret_f32(ptr addrspace(3) %ptr) nounwind {
-; VI-LABEL: lds_atomic_fadd_noret_f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v1, 4.0
-; VI-NEXT: s_mov_b32 m0, -1
-; VI-NEXT: ds_add_f32 v0, v1
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: lds_atomic_fadd_noret_f32:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v1, 4.0
-; GFX9-NEXT: ds_add_f32 v0, v1
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: lds_atomic_fadd_noret_f32:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v1, 4.0
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: ds_add_f32 v0, v1
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SE
-; GFX12-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX7-LABEL: lds_atomic_fadd_noret_f32:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: s_mov_b32 m0, -1
-; GFX7-NEXT: ds_read_b32 v1, v0
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: .LBB1_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_add_f32_e32 v2, 4.0, v1
-; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: v_mov_b32_e32 v1, v2
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB1_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: lds_atomic_fadd_noret_f32:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: s_mov_b32 m0, -1
-; GFX8-NEXT: ds_read_b32 v1, v0
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB1_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_add_f32_e32 v2, 4.0, v1
-; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v1, v2
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB1_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr addrspace(3) %ptr, float 4.0 seq_cst
- ret void
-}
-
-define amdgpu_kernel void @lds_ds_fadd(ptr addrspace(1) %out, ptr addrspace(3) %ptrf, i32 %idx) {
-; VI-LABEL: lds_ds_fadd:
-; VI: ; %bb.0:
-; VI-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x2c
-; VI-NEXT: s_mov_b64 s[4:5], exec
-; VI-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; VI-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: s_add_i32 s3, s3, 4
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: s_mov_b32 m0, -1
-; VI-NEXT: s_and_saveexec_b64 s[6:7], vcc
-; VI-NEXT: s_cbranch_execz .LBB2_2
-; VI-NEXT: ; %bb.1:
-; VI-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
-; VI-NEXT: s_lshl_b32 s8, s3, 3
-; VI-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
-; VI-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
-; VI-NEXT: v_mov_b32_e32 v2, s8
-; VI-NEXT: ds_add_rtn_f32 v1, v2, v1
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: .LBB2_2:
-; VI-NEXT: s_or_b64 exec, exec, s[6:7]
-; VI-NEXT: s_mov_b64 s[6:7], exec
-; VI-NEXT: v_readfirstlane_b32 s8, v1
-; VI-NEXT: v_mbcnt_lo_u32_b32 v1, s6, 0
-; VI-NEXT: v_mbcnt_hi_u32_b32 v1, s7, v1
-; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_cbranch_execz .LBB2_4
-; VI-NEXT: ; %bb.3:
-; VI-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
-; VI-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
-; VI-NEXT: s_lshl_b32 s3, s3, 4
-; VI-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
-; VI-NEXT: v_mov_b32_e32 v2, s3
-; VI-NEXT: ds_add_f32 v2, v1
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: .LBB2_4:
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; VI-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
-; VI-NEXT: s_mov_b64 s[4:5], exec
-; VI-NEXT: v_add_f32_e32 v2, s8, v0
-; VI-NEXT: v_bfrev_b32_e32 v1, 1
-; VI-NEXT: ; implicit-def: $vgpr0
-; VI-NEXT: .LBB2_5: ; %ComputeLoop
-; VI-NEXT: ; =>This Inner Loop Header: Depth=1
-; VI-NEXT: s_ff1_i32_b64 s3, s[4:5]
-; VI-NEXT: s_lshl_b64 s[6:7], 1, s3
-; VI-NEXT: v_readfirstlane_b32 s8, v1
-; VI-NEXT: v_readlane_b32 s9, v2, s3
-; VI-NEXT: s_mov_b32 m0, s3
-; VI-NEXT: s_andn2_b64 s[4:5], s[4:5], s[6:7]
-; VI-NEXT: v_writelane_b32 v0, s8, m0
-; VI-NEXT: s_cmp_lg_u64 s[4:5], 0
-; VI-NEXT: v_add_f32_e32 v1, s9, v1
-; VI-NEXT: s_cbranch_scc1 .LBB2_5
-; VI-NEXT: ; %bb.6: ; %ComputeEnd
-; VI-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
-; VI-NEXT: v_mbcnt_hi_u32_b32 v2, exec_hi, v2
-; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB2_8
-; VI-NEXT: ; %bb.7:
-; VI-NEXT: v_mov_b32_e32 v2, s2
-; VI-NEXT: s_mov_b32 m0, -1
-; VI-NEXT: ds_add_rtn_f32 v2, v2, v1
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: .LBB2_8:
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; VI-NEXT: v_readfirstlane_b32 s2, v2
-; VI-NEXT: v_add_f32_e32 v2, s2, v0
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v0, s0
-; VI-NEXT: v_mov_b32_e32 v1, s1
-; VI-NEXT: flat_store_dword v[0:1], v2
-; VI-NEXT: s_endpgm
-;
-; GFX908-LABEL: lds_ds_fadd:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x2c
-; GFX908-NEXT: s_mov_b64 s[4:5], exec
-; GFX908-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX908-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: s_add_i32 s3, s3, 4
-; GFX908-NEXT: ; implicit-def: $vgpr1
-; GFX908-NEXT: s_and_saveexec_b64 s[6:7], vcc
-; GFX908-NEXT: s_cbranch_execz .LBB2_2
-; GFX908-NEXT: ; %bb.1:
-; GFX908-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
-; GFX908-NEXT: s_lshl_b32 s8, s3, 3
-; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
-; GFX908-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
-; GFX908-NEXT: v_mov_b32_e32 v2, s8
-; GFX908-NEXT: ds_add_rtn_f32 v1, v2, v1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: .LBB2_2:
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_mov_b64 s[6:7], exec
-; GFX908-NEXT: v_readfirstlane_b32 s8, v1
-; GFX908-NEXT: v_mbcnt_lo_u32_b32 v1, s6, 0
-; GFX908-NEXT: v_mbcnt_hi_u32_b32 v1, s7, v1
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
-; GFX908-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX908-NEXT: s_cbranch_execz .LBB2_4
-; GFX908-NEXT: ; %bb.3:
-; GFX908-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
-; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
-; GFX908-NEXT: s_lshl_b32 s3, s3, 4
-; GFX908-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
-; GFX908-NEXT: v_mov_b32_e32 v2, s3
-; GFX908-NEXT: ds_add_f32 v2, v1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: .LBB2_4:
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX908-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
-; GFX908-NEXT: s_mov_b64 s[4:5], exec
-; GFX908-NEXT: v_add_f32_e32 v2, s8, v0
-; GFX908-NEXT: v_bfrev_b32_e32 v1, 1
-; GFX908-NEXT: ; implicit-def: $vgpr0
-; GFX908-NEXT: .LBB2_5: ; %ComputeLoop
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_ff1_i32_b64 s3, s[4:5]
-; GFX908-NEXT: s_lshl_b64 s[6:7], 1, s3
-; GFX908-NEXT: v_readfirstlane_b32 s8, v1
-; GFX908-NEXT: v_readlane_b32 s9, v2, s3
-; GFX908-NEXT: s_mov_b32 m0, s3
-; GFX908-NEXT: s_andn2_b64 s[4:5], s[4:5], s[6:7]
-; GFX908-NEXT: v_writelane_b32 v0, s8, m0
-; GFX908-NEXT: s_cmp_lg_u64 s[4:5], 0
-; GFX908-NEXT: v_add_f32_e32 v1, s9, v1
-; GFX908-NEXT: s_cbranch_scc1 .LBB2_5
-; GFX908-NEXT: ; %bb.6: ; %ComputeEnd
-; GFX908-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
-; GFX908-NEXT: v_mbcnt_hi_u32_b32 v2, exec_hi, v2
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
-; GFX908-NEXT: ; implicit-def: $vgpr2
-; GFX908-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX908-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execz .LBB2_8
-; GFX908-NEXT: ; %bb.7:
-; GFX908-NEXT: v_mov_b32_e32 v2, s2
-; GFX908-NEXT: ds_add_rtn_f32 v2, v2, v1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: .LBB2_8:
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX908-NEXT: v_readfirstlane_b32 s2, v2
-; GFX908-NEXT: v_mov_b32_e32 v1, 0
-; GFX908-NEXT: v_add_f32_e32 v0, s2, v0
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: global_store_dword v1, v0, s[0:1]
-; GFX908-NEXT: s_endpgm
-;
-; GFX90A-LABEL: lds_ds_fadd:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x2c
-; GFX90A-NEXT: s_mov_b64 s[4:5], exec
-; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: s_add_i32 s3, s3, 4
-; GFX90A-NEXT: ; implicit-def: $vgpr1
-; GFX90A-NEXT: s_and_saveexec_b64 s[6:7], vcc
-; GFX90A-NEXT: s_cbranch_execz .LBB2_2
-; GFX90A-NEXT: ; %bb.1:
-; GFX90A-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
-; GFX90A-NEXT: s_lshl_b32 s8, s3, 3
-; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
-; GFX90A-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
-; GFX90A-NEXT: v_mov_b32_e32 v2, s8
-; GFX90A-NEXT: ds_add_rtn_f32 v1, v2, v1
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: .LBB2_2:
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_mov_b64 s[6:7], exec
-; GFX90A-NEXT: v_readfirstlane_b32 s8, v1
-; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v1, s6, 0
-; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v1, s7, v1
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
-; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX90A-NEXT: s_cbranch_execz .LBB2_4
-; GFX90A-NEXT: ; %bb.3:
-; GFX90A-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
-; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
-; GFX90A-NEXT: s_lshl_b32 s3, s3, 4
-; GFX90A-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
-; GFX90A-NEXT: v_mov_b32_e32 v2, s3
-; GFX90A-NEXT: ds_add_f32 v2, v1
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: .LBB2_4:
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX90A-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
-; GFX90A-NEXT: s_mov_b64 s[4:5], exec
-; GFX90A-NEXT: v_add_f32_e32 v2, s8, v0
-; GFX90A-NEXT: v_bfrev_b32_e32 v1, 1
-; GFX90A-NEXT: ; implicit-def: $vgpr0
-; GFX90A-NEXT: .LBB2_5: ; %ComputeLoop
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_ff1_i32_b64 s3, s[4:5]
-; GFX90A-NEXT: s_lshl_b64 s[6:7], 1, s3
-; GFX90A-NEXT: v_readfirstlane_b32 s8, v1
-; GFX90A-NEXT: v_readlane_b32 s9, v2, s3
-; GFX90A-NEXT: s_mov_b32 m0, s3
-; GFX90A-NEXT: s_andn2_b64 s[4:5], s[4:5], s[6:7]
-; GFX90A-NEXT: v_writelane_b32 v0, s8, m0
-; GFX90A-NEXT: s_cmp_lg_u64 s[4:5], 0
-; GFX90A-NEXT: v_add_f32_e32 v1, s9, v1
-; GFX90A-NEXT: s_cbranch_scc1 .LBB2_5
-; GFX90A-NEXT: ; %bb.6: ; %ComputeEnd
-; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
-; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v2, exec_hi, v2
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
-; GFX90A-NEXT: ; implicit-def: $vgpr2
-; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX90A-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execz .LBB2_8
-; GFX90A-NEXT: ; %bb.7:
-; GFX90A-NEXT: v_mov_b32_e32 v2, s2
-; GFX90A-NEXT: ds_add_rtn_f32 v2, v2, v1
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: .LBB2_8:
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX90A-NEXT: v_readfirstlane_b32 s2, v2
-; GFX90A-NEXT: v_mov_b32_e32 v1, 0
-; GFX90A-NEXT: v_add_f32_e32 v0, s2, v0
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: global_store_dword v1, v0, s[0:1]
-; GFX90A-NEXT: s_endpgm
-;
-; GFX940-LABEL: lds_ds_fadd:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x2c
-; GFX940-NEXT: s_mov_b64 s[4:5], exec
-; GFX940-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX940-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: s_add_i32 s3, s3, 4
-; GFX940-NEXT: ; implicit-def: $vgpr1
-; GFX940-NEXT: s_and_saveexec_b64 s[6:7], vcc
-; GFX940-NEXT: s_cbranch_execz .LBB2_2
-; GFX940-NEXT: ; %bb.1:
-; GFX940-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
-; GFX940-NEXT: s_lshl_b32 s8, s3, 3
-; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
-; GFX940-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
-; GFX940-NEXT: v_mov_b32_e32 v2, s8
-; GFX940-NEXT: ds_add_rtn_f32 v1, v2, v1
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: .LBB2_2:
-; GFX940-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX940-NEXT: s_mov_b64 s[6:7], exec
-; GFX940-NEXT: v_readfirstlane_b32 s8, v1
-; GFX940-NEXT: v_mbcnt_lo_u32_b32 v1, s6, 0
-; GFX940-NEXT: v_mbcnt_hi_u32_b32 v1, s7, v1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
-; GFX940-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX940-NEXT: s_cbranch_execz .LBB2_4
-; GFX940-NEXT: ; %bb.3:
-; GFX940-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
-; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
-; GFX940-NEXT: s_lshl_b32 s3, s3, 4
-; GFX940-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
-; GFX940-NEXT: v_mov_b32_e32 v2, s3
-; GFX940-NEXT: ds_add_f32 v2, v1
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: .LBB2_4:
-; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX940-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
-; GFX940-NEXT: s_mov_b64 s[4:5], exec
-; GFX940-NEXT: v_add_f32_e32 v2, s8, v0
-; GFX940-NEXT: v_bfrev_b32_e32 v1, 1
-; GFX940-NEXT: ; implicit-def: $vgpr0
-; GFX940-NEXT: .LBB2_5: ; %ComputeLoop
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_ff1_i32_b64 s3, s[4:5]
-; GFX940-NEXT: s_lshl_b64 s[6:7], 1, s3
-; GFX940-NEXT: v_readfirstlane_b32 s8, v1
-; GFX940-NEXT: v_readlane_b32 s9, v2, s3
-; GFX940-NEXT: s_mov_b32 m0, s3
-; GFX940-NEXT: s_andn2_b64 s[4:5], s[4:5], s[6:7]
-; GFX940-NEXT: v_writelane_b32 v0, s8, m0
-; GFX940-NEXT: s_cmp_lg_u64 s[4:5], 0
-; GFX940-NEXT: v_add_f32_e32 v1, s9, v1
-; GFX940-NEXT: s_cbranch_scc1 .LBB2_5
-; GFX940-NEXT: ; %bb.6: ; %ComputeEnd
-; GFX940-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
-; GFX940-NEXT: v_mbcnt_hi_u32_b32 v2, exec_hi, v2
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
-; GFX940-NEXT: ; implicit-def: $vgpr2
-; GFX940-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX940-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX940-NEXT: s_cbranch_execz .LBB2_8
-; GFX940-NEXT: ; %bb.7:
-; GFX940-NEXT: v_mov_b32_e32 v2, s2
-; GFX940-NEXT: ds_add_rtn_f32 v2, v2, v1
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: .LBB2_8:
-; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX940-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX940-NEXT: v_readfirstlane_b32 s2, v2
-; GFX940-NEXT: v_mov_b32_e32 v1, 0
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_add_f32_e32 v0, s2, v0
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: global_store_dword v1, v0, s[0:1] sc0 sc1
-; GFX940-NEXT: s_endpgm
-;
-; GFX12-LABEL: lds_ds_fadd:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_load_b64 s[2:3], s[0:1], 0x2c
-; GFX12-NEXT: s_mov_b32 s5, exec_lo
-; GFX12-NEXT: s_mov_b32 s4, exec_lo
-; GFX12-NEXT: v_mbcnt_lo_u32_b32 v0, s5, 0
-; GFX12-NEXT: ; implicit-def: $vgpr1
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: s_add_co_i32 s3, s3, 4
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT: v_cmpx_eq_u32_e32 0, v0
-; GFX12-NEXT: s_cbranch_execz .LBB2_2
-; GFX12-NEXT: ; %bb.1:
-; GFX12-NEXT: s_bcnt1_i32_b32 s5, s5
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v1, s5
-; GFX12-NEXT: s_lshl_b32 s5, s3, 3
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
-; GFX12-NEXT: v_dual_mov_b32 v2, s5 :: v_dual_mul_f32 v1, 0x42280000, v1
-; GFX12-NEXT: ds_add_rtn_f32 v1, v2, v1
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SE
-; GFX12-NEXT: .LBB2_2:
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
-; GFX12-NEXT: s_mov_b32 s6, exec_lo
-; GFX12-NEXT: v_readfirstlane_b32 s5, v1
-; GFX12-NEXT: v_mbcnt_lo_u32_b32 v2, s6, 0
-; GFX12-NEXT: s_mov_b32 s4, exec_lo
-; GFX12-NEXT: v_cmpx_eq_u32_e32 0, v2
-; GFX12-NEXT: s_cbranch_execz .LBB2_4
-; GFX12-NEXT: ; %bb.3:
-; GFX12-NEXT: s_bcnt1_i32_b32 s6, s6
-; GFX12-NEXT: s_lshl_b32 s3, s3, 4
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT: v_dual_mov_b32 v2, s3 :: v_dual_mul_f32 v1, 0x42280000, v1
-; GFX12-NEXT: ds_add_f32 v2, v1
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SE
-; GFX12-NEXT: .LBB2_4:
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX12-NEXT: s_mov_b32 s4, exec_lo
-; GFX12-NEXT: s_brev_b32 s3, 1
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
-; GFX12-NEXT: v_add_f32_e32 v1, s5, v0
-; GFX12-NEXT: ; implicit-def: $vgpr0
-; GFX12-NEXT: .LBB2_5: ; %ComputeLoop
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_ctz_i32_b32 s5, s4
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
-; GFX12-NEXT: v_readlane_b32 s6, v1, s5
-; GFX12-NEXT: s_lshl_b32 s7, 1, s5
-; GFX12-NEXT: v_writelane_b32 v0, s3, s5
-; GFX12-NEXT: s_and_not1_b32 s4, s4, s7
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_cmp_lg_u32 s4, 0
-; GFX12-NEXT: s_add_f32 s3, s3, s6
-; GFX12-NEXT: s_cbranch_scc1 .LBB2_5
-; GFX12-NEXT: ; %bb.6: ; %ComputeEnd
-; GFX12-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
-; GFX12-NEXT: ; implicit-def: $vgpr1
-; GFX12-NEXT: s_and_saveexec_b32 s4, vcc_lo
-; GFX12-NEXT: s_xor_b32 s4, exec_lo, s4
-; GFX12-NEXT: s_cbranch_execz .LBB2_8
-; GFX12-NEXT: ; %bb.7:
-; GFX12-NEXT: v_dual_mov_b32 v1, s2 :: v_dual_mov_b32 v2, s3
-; GFX12-NEXT: ds_add_rtn_f32 v1, v1, v2
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SE
-; GFX12-NEXT: .LBB2_8:
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX12-NEXT: s_load_b64 s[0:1], s[0:1], 0x24
-; GFX12-NEXT: v_readfirstlane_b32 s2, v1
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_add_f32 v0, s2, v0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_store_b32 v1, v0, s[0:1]
-; GFX12-NEXT: s_nop 0
-; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX12-NEXT: s_endpgm
-;
-; GFX7-LABEL: lds_ds_fadd:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0xb
-; GFX7-NEXT: s_mov_b64 s[6:7], exec
-; GFX7-NEXT: v_mbcnt_lo_u32_b32_e64 v0, s6, 0
-; GFX7-NEXT: v_mbcnt_hi_u32_b32_e32 v0, s7, v0
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: s_add_i32 s3, s3, 4
-; GFX7-NEXT: ; implicit-def: $vgpr1
-; GFX7-NEXT: s_mov_b32 m0, -1
-; GFX7-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX7-NEXT: s_cbranch_execz .LBB2_4
-; GFX7-NEXT: ; %bb.1:
-; GFX7-NEXT: s_lshl_b32 s8, s3, 3
-; GFX7-NEXT: v_mov_b32_e32 v2, s8
-; GFX7-NEXT: ds_read_b32 v1, v2
-; GFX7-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
-; GFX7-NEXT: v_cvt_f32_ubyte0_e32 v3, s6
-; GFX7-NEXT: v_mul_f32_e32 v3, 0x42280000, v3
-; GFX7-NEXT: s_mov_b64 s[6:7], 0
-; GFX7-NEXT: .LBB2_2: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_mov_b32_e32 v4, v1
-; GFX7-NEXT: v_add_f32_e32 v1, v4, v3
-; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v2, v4, v1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v4
-; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX7-NEXT: s_cbranch_execnz .LBB2_2
-; GFX7-NEXT: ; %bb.3: ; %Flow15
-; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX7-NEXT: .LBB2_4: ; %Flow16
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_mov_b64 s[6:7], exec
-; GFX7-NEXT: v_readfirstlane_b32 s8, v1
-; GFX7-NEXT: v_mbcnt_lo_u32_b32_e64 v1, s6, 0
-; GFX7-NEXT: v_mbcnt_hi_u32_b32_e32 v1, s7, v1
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
-; GFX7-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX7-NEXT: s_cbranch_execz .LBB2_7
-; GFX7-NEXT: ; %bb.5:
-; GFX7-NEXT: s_lshl_b32 s3, s3, 4
-; GFX7-NEXT: v_mov_b32_e32 v1, s3
-; GFX7-NEXT: ds_read_b32 v3, v1
-; GFX7-NEXT: s_bcnt1_i32_b64 s3, s[6:7]
-; GFX7-NEXT: v_cvt_f32_ubyte0_e32 v2, s3
-; GFX7-NEXT: v_mul_f32_e32 v2, 0x42280000, v2
-; GFX7-NEXT: s_mov_b64 s[6:7], 0
-; GFX7-NEXT: .LBB2_6: ; %atomicrmw.start2
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_add_f32_e32 v4, v3, v2
-; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
-; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX7-NEXT: v_mov_b32_e32 v3, v4
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX7-NEXT: s_cbranch_execnz .LBB2_6
-; GFX7-NEXT: .LBB2_7: ; %Flow14
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: v_mov_b32_e32 v2, s2
-; GFX7-NEXT: ds_read_b32 v1, v2
-; GFX7-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX7-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
-; GFX7-NEXT: v_add_f32_e32 v0, s8, v0
-; GFX7-NEXT: s_mov_b64 s[2:3], 0
-; GFX7-NEXT: .LBB2_8: ; %atomicrmw.start8
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_mov_b32_e32 v3, v1
-; GFX7-NEXT: v_add_f32_e32 v1, v3, v0
-; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v2, v3, v1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v3
-; GFX7-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX7-NEXT: s_cbranch_execnz .LBB2_8
-; GFX7-NEXT: ; %bb.9: ; %atomicrmw.end7
-; GFX7-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX7-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
-; GFX7-NEXT: s_mov_b32 s3, 0xf000
-; GFX7-NEXT: s_mov_b32 s2, -1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: buffer_store_dword v1, off, s[0:3], 0
-; GFX7-NEXT: s_endpgm
-;
-; GFX8-LABEL: lds_ds_fadd:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0xb
-; GFX8-NEXT: s_mov_b64 s[6:7], exec
-; GFX8-NEXT: v_mbcnt_lo_u32_b32_e64 v0, s6, 0
-; GFX8-NEXT: v_mbcnt_hi_u32_b32_e32 v0, s7, v0
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: s_add_i32 s3, s3, 4
-; GFX8-NEXT: ; implicit-def: $vgpr1
-; GFX8-NEXT: s_mov_b32 m0, -1
-; GFX8-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX8-NEXT: s_cbranch_execz .LBB2_4
-; GFX8-NEXT: ; %bb.1:
-; GFX8-NEXT: s_lshl_b32 s8, s3, 3
-; GFX8-NEXT: v_mov_b32_e32 v2, s8
-; GFX8-NEXT: ds_read_b32 v1, v2
-; GFX8-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
-; GFX8-NEXT: v_cvt_f32_ubyte0_e32 v3, s6
-; GFX8-NEXT: v_mul_f32_e32 v3, 0x42280000, v3
-; GFX8-NEXT: s_mov_b64 s[6:7], 0
-; GFX8-NEXT: .LBB2_2: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_mov_b32_e32 v4, v1
-; GFX8-NEXT: v_add_f32_e32 v1, v4, v3
-; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v2, v4, v1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v4
-; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX8-NEXT: s_cbranch_execnz .LBB2_2
-; GFX8-NEXT: ; %bb.3: ; %Flow17
-; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX8-NEXT: .LBB2_4: ; %Flow18
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_mov_b64 s[6:7], exec
-; GFX8-NEXT: v_readfirstlane_b32 s8, v1
-; GFX8-NEXT: v_mbcnt_lo_u32_b32_e64 v1, s6, 0
-; GFX8-NEXT: v_mbcnt_hi_u32_b32_e32 v1, s7, v1
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
-; GFX8-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX8-NEXT: s_cbranch_execz .LBB2_7
-; GFX8-NEXT: ; %bb.5:
-; GFX8-NEXT: s_lshl_b32 s3, s3, 4
-; GFX8-NEXT: v_mov_b32_e32 v1, s3
-; GFX8-NEXT: ds_read_b32 v3, v1
-; GFX8-NEXT: s_bcnt1_i32_b64 s3, s[6:7]
-; GFX8-NEXT: v_cvt_f32_ubyte0_e32 v2, s3
-; GFX8-NEXT: v_mul_f32_e32 v2, 0x42280000, v2
-; GFX8-NEXT: s_mov_b64 s[6:7], 0
-; GFX8-NEXT: .LBB2_6: ; %atomicrmw.start2
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_add_f32_e32 v4, v3, v2
-; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
-; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX8-NEXT: v_mov_b32_e32 v3, v4
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX8-NEXT: s_cbranch_execnz .LBB2_6
-; GFX8-NEXT: .LBB2_7: ; %Flow16
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v2, s2
-; GFX8-NEXT: ds_read_b32 v1, v2
-; GFX8-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX8-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
-; GFX8-NEXT: v_add_f32_e32 v0, s8, v0
-; GFX8-NEXT: s_mov_b64 s[2:3], 0
-; GFX8-NEXT: .LBB2_8: ; %atomicrmw.start8
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_mov_b32_e32 v3, v1
-; GFX8-NEXT: v_add_f32_e32 v1, v3, v0
-; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v2, v3, v1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v3
-; GFX8-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX8-NEXT: s_cbranch_execnz .LBB2_8
-; GFX8-NEXT: ; %bb.9: ; %atomicrmw.end7
-; GFX8-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX8-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
-; GFX8-NEXT: s_mov_b32 s3, 0xf000
-; GFX8-NEXT: s_mov_b32 s2, -1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: buffer_store_dword v1, off, s[0:3], 0
-; GFX8-NEXT: s_endpgm
- %idx.add = add nuw i32 %idx, 4
- %shl0 = shl i32 %idx.add, 3
- %shl1 = shl i32 %idx.add, 4
- %ptr0 = inttoptr i32 %shl0 to ptr addrspace(3)
- %ptr1 = inttoptr i32 %shl1 to ptr addrspace(3)
- %a1 = atomicrmw fadd ptr addrspace(3) %ptr0, float 4.2e+1 seq_cst
- %a2 = atomicrmw fadd ptr addrspace(3) %ptr1, float 4.2e+1 seq_cst
- %a3 = atomicrmw fadd ptr addrspace(3) %ptrf, float %a1 seq_cst
- store float %a3, ptr addrspace(1) %out
- ret void
-}
-
-define amdgpu_kernel void @lds_ds_fadd_one_as(ptr addrspace(1) %out, ptr addrspace(3) %ptrf, i32 %idx) {
-; VI-LABEL: lds_ds_fadd_one_as:
-; VI: ; %bb.0:
-; VI-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x2c
-; VI-NEXT: s_mov_b64 s[4:5], exec
-; VI-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; VI-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: s_add_i32 s3, s3, 4
-; VI-NEXT: ; implicit-def: $vgpr1
-; VI-NEXT: s_mov_b32 m0, -1
-; VI-NEXT: s_and_saveexec_b64 s[6:7], vcc
-; VI-NEXT: s_cbranch_execz .LBB3_2
-; VI-NEXT: ; %bb.1:
-; VI-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
-; VI-NEXT: s_lshl_b32 s8, s3, 3
-; VI-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
-; VI-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
-; VI-NEXT: v_mov_b32_e32 v2, s8
-; VI-NEXT: ds_add_rtn_f32 v1, v2, v1
-; VI-NEXT: .LBB3_2:
-; VI-NEXT: s_or_b64 exec, exec, s[6:7]
-; VI-NEXT: s_mov_b64 s[6:7], exec
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_readfirstlane_b32 s8, v1
-; VI-NEXT: v_mbcnt_lo_u32_b32 v1, s6, 0
-; VI-NEXT: v_mbcnt_hi_u32_b32 v1, s7, v1
-; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_cbranch_execz .LBB3_4
-; VI-NEXT: ; %bb.3:
-; VI-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
-; VI-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
-; VI-NEXT: s_lshl_b32 s3, s3, 4
-; VI-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
-; VI-NEXT: v_mov_b32_e32 v2, s3
-; VI-NEXT: ds_add_f32 v2, v1
-; VI-NEXT: .LBB3_4:
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; VI-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
-; VI-NEXT: s_mov_b64 s[4:5], exec
-; VI-NEXT: v_add_f32_e32 v2, s8, v0
-; VI-NEXT: v_bfrev_b32_e32 v1, 1
-; VI-NEXT: ; implicit-def: $vgpr0
-; VI-NEXT: .LBB3_5: ; %ComputeLoop
-; VI-NEXT: ; =>This Inner Loop Header: Depth=1
-; VI-NEXT: s_ff1_i32_b64 s3, s[4:5]
-; VI-NEXT: s_lshl_b64 s[6:7], 1, s3
-; VI-NEXT: v_readfirstlane_b32 s8, v1
-; VI-NEXT: v_readlane_b32 s9, v2, s3
-; VI-NEXT: s_mov_b32 m0, s3
-; VI-NEXT: s_andn2_b64 s[4:5], s[4:5], s[6:7]
-; VI-NEXT: v_writelane_b32 v0, s8, m0
-; VI-NEXT: s_cmp_lg_u64 s[4:5], 0
-; VI-NEXT: v_add_f32_e32 v1, s9, v1
-; VI-NEXT: s_cbranch_scc1 .LBB3_5
-; VI-NEXT: ; %bb.6: ; %ComputeEnd
-; VI-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
-; VI-NEXT: v_mbcnt_hi_u32_b32 v2, exec_hi, v2
-; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
-; VI-NEXT: ; implicit-def: $vgpr2
-; VI-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; VI-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; VI-NEXT: s_cbranch_execz .LBB3_8
-; VI-NEXT: ; %bb.7:
-; VI-NEXT: v_mov_b32_e32 v2, s2
-; VI-NEXT: s_mov_b32 m0, -1
-; VI-NEXT: ds_add_rtn_f32 v2, v2, v1
-; VI-NEXT: .LBB3_8:
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_readfirstlane_b32 s2, v2
-; VI-NEXT: v_add_f32_e32 v2, s2, v0
-; VI-NEXT: v_mov_b32_e32 v0, s0
-; VI-NEXT: v_mov_b32_e32 v1, s1
-; VI-NEXT: flat_store_dword v[0:1], v2
-; VI-NEXT: s_endpgm
-;
-; GFX908-LABEL: lds_ds_fadd_one_as:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x2c
-; GFX908-NEXT: s_mov_b64 s[4:5], exec
-; GFX908-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX908-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: s_add_i32 s3, s3, 4
-; GFX908-NEXT: ; implicit-def: $vgpr1
-; GFX908-NEXT: s_and_saveexec_b64 s[6:7], vcc
-; GFX908-NEXT: s_cbranch_execz .LBB3_2
-; GFX908-NEXT: ; %bb.1:
-; GFX908-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
-; GFX908-NEXT: s_lshl_b32 s8, s3, 3
-; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
-; GFX908-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
-; GFX908-NEXT: v_mov_b32_e32 v2, s8
-; GFX908-NEXT: ds_add_rtn_f32 v1, v2, v1
-; GFX908-NEXT: .LBB3_2:
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_mov_b64 s[6:7], exec
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_readfirstlane_b32 s8, v1
-; GFX908-NEXT: v_mbcnt_lo_u32_b32 v1, s6, 0
-; GFX908-NEXT: v_mbcnt_hi_u32_b32 v1, s7, v1
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
-; GFX908-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX908-NEXT: s_cbranch_execz .LBB3_4
-; GFX908-NEXT: ; %bb.3:
-; GFX908-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
-; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
-; GFX908-NEXT: s_lshl_b32 s3, s3, 4
-; GFX908-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
-; GFX908-NEXT: v_mov_b32_e32 v2, s3
-; GFX908-NEXT: ds_add_f32 v2, v1
-; GFX908-NEXT: .LBB3_4:
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX908-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
-; GFX908-NEXT: s_mov_b64 s[4:5], exec
-; GFX908-NEXT: v_add_f32_e32 v2, s8, v0
-; GFX908-NEXT: v_bfrev_b32_e32 v1, 1
-; GFX908-NEXT: ; implicit-def: $vgpr0
-; GFX908-NEXT: .LBB3_5: ; %ComputeLoop
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_ff1_i32_b64 s3, s[4:5]
-; GFX908-NEXT: s_lshl_b64 s[6:7], 1, s3
-; GFX908-NEXT: v_readfirstlane_b32 s8, v1
-; GFX908-NEXT: v_readlane_b32 s9, v2, s3
-; GFX908-NEXT: s_mov_b32 m0, s3
-; GFX908-NEXT: s_andn2_b64 s[4:5], s[4:5], s[6:7]
-; GFX908-NEXT: v_writelane_b32 v0, s8, m0
-; GFX908-NEXT: s_cmp_lg_u64 s[4:5], 0
-; GFX908-NEXT: v_add_f32_e32 v1, s9, v1
-; GFX908-NEXT: s_cbranch_scc1 .LBB3_5
-; GFX908-NEXT: ; %bb.6: ; %ComputeEnd
-; GFX908-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
-; GFX908-NEXT: v_mbcnt_hi_u32_b32 v2, exec_hi, v2
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
-; GFX908-NEXT: ; implicit-def: $vgpr2
-; GFX908-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX908-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execz .LBB3_8
-; GFX908-NEXT: ; %bb.7:
-; GFX908-NEXT: v_mov_b32_e32 v2, s2
-; GFX908-NEXT: ds_add_rtn_f32 v2, v2, v1
-; GFX908-NEXT: .LBB3_8:
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_readfirstlane_b32 s2, v2
-; GFX908-NEXT: v_mov_b32_e32 v1, 0
-; GFX908-NEXT: v_add_f32_e32 v0, s2, v0
-; GFX908-NEXT: global_store_dword v1, v0, s[0:1]
-; GFX908-NEXT: s_endpgm
-;
-; GFX90A-LABEL: lds_ds_fadd_one_as:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x2c
-; GFX90A-NEXT: s_mov_b64 s[4:5], exec
-; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: s_add_i32 s3, s3, 4
-; GFX90A-NEXT: ; implicit-def: $vgpr1
-; GFX90A-NEXT: s_and_saveexec_b64 s[6:7], vcc
-; GFX90A-NEXT: s_cbranch_execz .LBB3_2
-; GFX90A-NEXT: ; %bb.1:
-; GFX90A-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
-; GFX90A-NEXT: s_lshl_b32 s8, s3, 3
-; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
-; GFX90A-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
-; GFX90A-NEXT: v_mov_b32_e32 v2, s8
-; GFX90A-NEXT: ds_add_rtn_f32 v1, v2, v1
-; GFX90A-NEXT: .LBB3_2:
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_mov_b64 s[6:7], exec
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_readfirstlane_b32 s8, v1
-; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v1, s6, 0
-; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v1, s7, v1
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
-; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX90A-NEXT: s_cbranch_execz .LBB3_4
-; GFX90A-NEXT: ; %bb.3:
-; GFX90A-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
-; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
-; GFX90A-NEXT: s_lshl_b32 s3, s3, 4
-; GFX90A-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
-; GFX90A-NEXT: v_mov_b32_e32 v2, s3
-; GFX90A-NEXT: ds_add_f32 v2, v1
-; GFX90A-NEXT: .LBB3_4:
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX90A-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
-; GFX90A-NEXT: s_mov_b64 s[4:5], exec
-; GFX90A-NEXT: v_add_f32_e32 v2, s8, v0
-; GFX90A-NEXT: v_bfrev_b32_e32 v1, 1
-; GFX90A-NEXT: ; implicit-def: $vgpr0
-; GFX90A-NEXT: .LBB3_5: ; %ComputeLoop
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_ff1_i32_b64 s3, s[4:5]
-; GFX90A-NEXT: s_lshl_b64 s[6:7], 1, s3
-; GFX90A-NEXT: v_readfirstlane_b32 s8, v1
-; GFX90A-NEXT: v_readlane_b32 s9, v2, s3
-; GFX90A-NEXT: s_mov_b32 m0, s3
-; GFX90A-NEXT: s_andn2_b64 s[4:5], s[4:5], s[6:7]
-; GFX90A-NEXT: v_writelane_b32 v0, s8, m0
-; GFX90A-NEXT: s_cmp_lg_u64 s[4:5], 0
-; GFX90A-NEXT: v_add_f32_e32 v1, s9, v1
-; GFX90A-NEXT: s_cbranch_scc1 .LBB3_5
-; GFX90A-NEXT: ; %bb.6: ; %ComputeEnd
-; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
-; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v2, exec_hi, v2
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
-; GFX90A-NEXT: ; implicit-def: $vgpr2
-; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX90A-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execz .LBB3_8
-; GFX90A-NEXT: ; %bb.7:
-; GFX90A-NEXT: v_mov_b32_e32 v2, s2
-; GFX90A-NEXT: ds_add_rtn_f32 v2, v2, v1
-; GFX90A-NEXT: .LBB3_8:
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_readfirstlane_b32 s2, v2
-; GFX90A-NEXT: v_mov_b32_e32 v1, 0
-; GFX90A-NEXT: v_add_f32_e32 v0, s2, v0
-; GFX90A-NEXT: global_store_dword v1, v0, s[0:1]
-; GFX90A-NEXT: s_endpgm
-;
-; GFX940-LABEL: lds_ds_fadd_one_as:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x2c
-; GFX940-NEXT: s_mov_b64 s[4:5], exec
-; GFX940-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX940-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: s_add_i32 s3, s3, 4
-; GFX940-NEXT: ; implicit-def: $vgpr1
-; GFX940-NEXT: s_and_saveexec_b64 s[6:7], vcc
-; GFX940-NEXT: s_cbranch_execz .LBB3_2
-; GFX940-NEXT: ; %bb.1:
-; GFX940-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
-; GFX940-NEXT: s_lshl_b32 s8, s3, 3
-; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
-; GFX940-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
-; GFX940-NEXT: v_mov_b32_e32 v2, s8
-; GFX940-NEXT: ds_add_rtn_f32 v1, v2, v1
-; GFX940-NEXT: .LBB3_2:
-; GFX940-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX940-NEXT: s_mov_b64 s[6:7], exec
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_readfirstlane_b32 s8, v1
-; GFX940-NEXT: v_mbcnt_lo_u32_b32 v1, s6, 0
-; GFX940-NEXT: v_mbcnt_hi_u32_b32 v1, s7, v1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
-; GFX940-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX940-NEXT: s_cbranch_execz .LBB3_4
-; GFX940-NEXT: ; %bb.3:
-; GFX940-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
-; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
-; GFX940-NEXT: s_lshl_b32 s3, s3, 4
-; GFX940-NEXT: v_mul_f32_e32 v1, 0x42280000, v1
-; GFX940-NEXT: v_mov_b32_e32 v2, s3
-; GFX940-NEXT: ds_add_f32 v2, v1
-; GFX940-NEXT: .LBB3_4:
-; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX940-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
-; GFX940-NEXT: s_mov_b64 s[4:5], exec
-; GFX940-NEXT: v_add_f32_e32 v2, s8, v0
-; GFX940-NEXT: v_bfrev_b32_e32 v1, 1
-; GFX940-NEXT: ; implicit-def: $vgpr0
-; GFX940-NEXT: .LBB3_5: ; %ComputeLoop
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_ff1_i32_b64 s3, s[4:5]
-; GFX940-NEXT: s_lshl_b64 s[6:7], 1, s3
-; GFX940-NEXT: v_readfirstlane_b32 s8, v1
-; GFX940-NEXT: v_readlane_b32 s9, v2, s3
-; GFX940-NEXT: s_mov_b32 m0, s3
-; GFX940-NEXT: s_andn2_b64 s[4:5], s[4:5], s[6:7]
-; GFX940-NEXT: v_writelane_b32 v0, s8, m0
-; GFX940-NEXT: s_cmp_lg_u64 s[4:5], 0
-; GFX940-NEXT: v_add_f32_e32 v1, s9, v1
-; GFX940-NEXT: s_cbranch_scc1 .LBB3_5
-; GFX940-NEXT: ; %bb.6: ; %ComputeEnd
-; GFX940-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
-; GFX940-NEXT: v_mbcnt_hi_u32_b32 v2, exec_hi, v2
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
-; GFX940-NEXT: ; implicit-def: $vgpr2
-; GFX940-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX940-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
-; GFX940-NEXT: s_cbranch_execz .LBB3_8
-; GFX940-NEXT: ; %bb.7:
-; GFX940-NEXT: v_mov_b32_e32 v2, s2
-; GFX940-NEXT: ds_add_rtn_f32 v2, v2, v1
-; GFX940-NEXT: .LBB3_8:
-; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX940-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_readfirstlane_b32 s2, v2
-; GFX940-NEXT: v_mov_b32_e32 v1, 0
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_add_f32_e32 v0, s2, v0
-; GFX940-NEXT: global_store_dword v1, v0, s[0:1] sc0 sc1
-; GFX940-NEXT: s_endpgm
-;
-; GFX12-LABEL: lds_ds_fadd_one_as:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_load_b64 s[2:3], s[0:1], 0x2c
-; GFX12-NEXT: s_mov_b32 s5, exec_lo
-; GFX12-NEXT: s_mov_b32 s4, exec_lo
-; GFX12-NEXT: v_mbcnt_lo_u32_b32 v0, s5, 0
-; GFX12-NEXT: ; implicit-def: $vgpr1
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: s_add_co_i32 s3, s3, 4
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT: v_cmpx_eq_u32_e32 0, v0
-; GFX12-NEXT: s_cbranch_execz .LBB3_2
-; GFX12-NEXT: ; %bb.1:
-; GFX12-NEXT: s_bcnt1_i32_b32 s5, s5
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v1, s5
-; GFX12-NEXT: s_lshl_b32 s5, s3, 3
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
-; GFX12-NEXT: v_dual_mov_b32 v2, s5 :: v_dual_mul_f32 v1, 0x42280000, v1
-; GFX12-NEXT: ds_add_rtn_f32 v1, v2, v1
-; GFX12-NEXT: .LBB3_2:
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(VALU_DEP_1)
-; GFX12-NEXT: s_mov_b32 s6, exec_lo
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: v_readfirstlane_b32 s5, v1
-; GFX12-NEXT: v_mbcnt_lo_u32_b32 v2, s6, 0
-; GFX12-NEXT: s_mov_b32 s4, exec_lo
-; GFX12-NEXT: v_cmpx_eq_u32_e32 0, v2
-; GFX12-NEXT: s_cbranch_execz .LBB3_4
-; GFX12-NEXT: ; %bb.3:
-; GFX12-NEXT: s_bcnt1_i32_b32 s6, s6
-; GFX12-NEXT: s_lshl_b32 s3, s3, 4
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v1, s6
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT: v_dual_mov_b32 v2, s3 :: v_dual_mul_f32 v1, 0x42280000, v1
-; GFX12-NEXT: ds_add_f32 v2, v1
-; GFX12-NEXT: .LBB3_4:
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX12-NEXT: s_mov_b32 s4, exec_lo
-; GFX12-NEXT: s_brev_b32 s3, 1
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
-; GFX12-NEXT: v_add_f32_e32 v1, s5, v0
-; GFX12-NEXT: ; implicit-def: $vgpr0
-; GFX12-NEXT: .LBB3_5: ; %ComputeLoop
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_ctz_i32_b32 s5, s4
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
-; GFX12-NEXT: v_readlane_b32 s6, v1, s5
-; GFX12-NEXT: s_lshl_b32 s7, 1, s5
-; GFX12-NEXT: v_writelane_b32 v0, s3, s5
-; GFX12-NEXT: s_and_not1_b32 s4, s4, s7
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_cmp_lg_u32 s4, 0
-; GFX12-NEXT: s_add_f32 s3, s3, s6
-; GFX12-NEXT: s_cbranch_scc1 .LBB3_5
-; GFX12-NEXT: ; %bb.6: ; %ComputeEnd
-; GFX12-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
-; GFX12-NEXT: ; implicit-def: $vgpr1
-; GFX12-NEXT: s_and_saveexec_b32 s4, vcc_lo
-; GFX12-NEXT: s_xor_b32 s4, exec_lo, s4
-; GFX12-NEXT: s_cbranch_execz .LBB3_8
-; GFX12-NEXT: ; %bb.7:
-; GFX12-NEXT: v_dual_mov_b32 v1, s2 :: v_dual_mov_b32 v2, s3
-; GFX12-NEXT: ds_add_rtn_f32 v1, v1, v2
-; GFX12-NEXT: .LBB3_8:
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX12-NEXT: s_load_b64 s[0:1], s[0:1], 0x24
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: v_readfirstlane_b32 s2, v1
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_add_f32 v0, s2, v0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_store_b32 v1, v0, s[0:1]
-; GFX12-NEXT: s_nop 0
-; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX12-NEXT: s_endpgm
-;
-; GFX7-LABEL: lds_ds_fadd_one_as:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0xb
-; GFX7-NEXT: s_mov_b64 s[6:7], exec
-; GFX7-NEXT: v_mbcnt_lo_u32_b32_e64 v0, s6, 0
-; GFX7-NEXT: v_mbcnt_hi_u32_b32_e32 v0, s7, v0
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: s_add_i32 s3, s3, 4
-; GFX7-NEXT: ; implicit-def: $vgpr1
-; GFX7-NEXT: s_mov_b32 m0, -1
-; GFX7-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX7-NEXT: s_cbranch_execz .LBB3_4
-; GFX7-NEXT: ; %bb.1:
-; GFX7-NEXT: s_lshl_b32 s8, s3, 3
-; GFX7-NEXT: v_mov_b32_e32 v2, s8
-; GFX7-NEXT: ds_read_b32 v1, v2
-; GFX7-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
-; GFX7-NEXT: v_cvt_f32_ubyte0_e32 v3, s6
-; GFX7-NEXT: v_mul_f32_e32 v3, 0x42280000, v3
-; GFX7-NEXT: s_mov_b64 s[6:7], 0
-; GFX7-NEXT: .LBB3_2: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_mov_b32_e32 v4, v1
-; GFX7-NEXT: v_add_f32_e32 v1, v4, v3
-; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v2, v4, v1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v4
-; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX7-NEXT: s_cbranch_execnz .LBB3_2
-; GFX7-NEXT: ; %bb.3: ; %Flow15
-; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX7-NEXT: .LBB3_4: ; %Flow16
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_mov_b64 s[6:7], exec
-; GFX7-NEXT: v_readfirstlane_b32 s8, v1
-; GFX7-NEXT: v_mbcnt_lo_u32_b32_e64 v1, s6, 0
-; GFX7-NEXT: v_mbcnt_hi_u32_b32_e32 v1, s7, v1
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
-; GFX7-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX7-NEXT: s_cbranch_execz .LBB3_7
-; GFX7-NEXT: ; %bb.5:
-; GFX7-NEXT: s_lshl_b32 s3, s3, 4
-; GFX7-NEXT: v_mov_b32_e32 v1, s3
-; GFX7-NEXT: ds_read_b32 v3, v1
-; GFX7-NEXT: s_bcnt1_i32_b64 s3, s[6:7]
-; GFX7-NEXT: v_cvt_f32_ubyte0_e32 v2, s3
-; GFX7-NEXT: v_mul_f32_e32 v2, 0x42280000, v2
-; GFX7-NEXT: s_mov_b64 s[6:7], 0
-; GFX7-NEXT: .LBB3_6: ; %atomicrmw.start2
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_add_f32_e32 v4, v3, v2
-; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
-; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX7-NEXT: v_mov_b32_e32 v3, v4
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX7-NEXT: s_cbranch_execnz .LBB3_6
-; GFX7-NEXT: .LBB3_7: ; %Flow14
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: v_mov_b32_e32 v2, s2
-; GFX7-NEXT: ds_read_b32 v1, v2
-; GFX7-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX7-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
-; GFX7-NEXT: v_add_f32_e32 v0, s8, v0
-; GFX7-NEXT: s_mov_b64 s[2:3], 0
-; GFX7-NEXT: .LBB3_8: ; %atomicrmw.start8
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_mov_b32_e32 v3, v1
-; GFX7-NEXT: v_add_f32_e32 v1, v3, v0
-; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v2, v3, v1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v3
-; GFX7-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX7-NEXT: s_cbranch_execnz .LBB3_8
-; GFX7-NEXT: ; %bb.9: ; %atomicrmw.end7
-; GFX7-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX7-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
-; GFX7-NEXT: s_mov_b32 s3, 0xf000
-; GFX7-NEXT: s_mov_b32 s2, -1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: buffer_store_dword v1, off, s[0:3], 0
-; GFX7-NEXT: s_endpgm
-;
-; GFX8-LABEL: lds_ds_fadd_one_as:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0xb
-; GFX8-NEXT: s_mov_b64 s[6:7], exec
-; GFX8-NEXT: v_mbcnt_lo_u32_b32_e64 v0, s6, 0
-; GFX8-NEXT: v_mbcnt_hi_u32_b32_e32 v0, s7, v0
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: s_add_i32 s3, s3, 4
-; GFX8-NEXT: ; implicit-def: $vgpr1
-; GFX8-NEXT: s_mov_b32 m0, -1
-; GFX8-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX8-NEXT: s_cbranch_execz .LBB3_4
-; GFX8-NEXT: ; %bb.1:
-; GFX8-NEXT: s_lshl_b32 s8, s3, 3
-; GFX8-NEXT: v_mov_b32_e32 v2, s8
-; GFX8-NEXT: ds_read_b32 v1, v2
-; GFX8-NEXT: s_bcnt1_i32_b64 s6, s[6:7]
-; GFX8-NEXT: v_cvt_f32_ubyte0_e32 v3, s6
-; GFX8-NEXT: v_mul_f32_e32 v3, 0x42280000, v3
-; GFX8-NEXT: s_mov_b64 s[6:7], 0
-; GFX8-NEXT: .LBB3_2: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_mov_b32_e32 v4, v1
-; GFX8-NEXT: v_add_f32_e32 v1, v4, v3
-; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v2, v4, v1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v4
-; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX8-NEXT: s_cbranch_execnz .LBB3_2
-; GFX8-NEXT: ; %bb.3: ; %Flow17
-; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX8-NEXT: .LBB3_4: ; %Flow18
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_mov_b64 s[6:7], exec
-; GFX8-NEXT: v_readfirstlane_b32 s8, v1
-; GFX8-NEXT: v_mbcnt_lo_u32_b32_e64 v1, s6, 0
-; GFX8-NEXT: v_mbcnt_hi_u32_b32_e32 v1, s7, v1
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
-; GFX8-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX8-NEXT: s_cbranch_execz .LBB3_7
-; GFX8-NEXT: ; %bb.5:
-; GFX8-NEXT: s_lshl_b32 s3, s3, 4
-; GFX8-NEXT: v_mov_b32_e32 v1, s3
-; GFX8-NEXT: ds_read_b32 v3, v1
-; GFX8-NEXT: s_bcnt1_i32_b64 s3, s[6:7]
-; GFX8-NEXT: v_cvt_f32_ubyte0_e32 v2, s3
-; GFX8-NEXT: v_mul_f32_e32 v2, 0x42280000, v2
-; GFX8-NEXT: s_mov_b64 s[6:7], 0
-; GFX8-NEXT: .LBB3_6: ; %atomicrmw.start2
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_add_f32_e32 v4, v3, v2
-; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
-; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX8-NEXT: v_mov_b32_e32 v3, v4
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX8-NEXT: s_cbranch_execnz .LBB3_6
-; GFX8-NEXT: .LBB3_7: ; %Flow16
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v2, s2
-; GFX8-NEXT: ds_read_b32 v1, v2
-; GFX8-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX8-NEXT: v_mul_f32_e32 v0, 0x42280000, v0
-; GFX8-NEXT: v_add_f32_e32 v0, s8, v0
-; GFX8-NEXT: s_mov_b64 s[2:3], 0
-; GFX8-NEXT: .LBB3_8: ; %atomicrmw.start8
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_mov_b32_e32 v3, v1
-; GFX8-NEXT: v_add_f32_e32 v1, v3, v0
-; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v2, v3, v1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v3
-; GFX8-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX8-NEXT: s_cbranch_execnz .LBB3_8
-; GFX8-NEXT: ; %bb.9: ; %atomicrmw.end7
-; GFX8-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX8-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
-; GFX8-NEXT: s_mov_b32 s3, 0xf000
-; GFX8-NEXT: s_mov_b32 s2, -1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: buffer_store_dword v1, off, s[0:3], 0
-; GFX8-NEXT: s_endpgm
- %idx.add = add nuw i32 %idx, 4
- %shl0 = shl i32 %idx.add, 3
- %shl1 = shl i32 %idx.add, 4
- %ptr0 = inttoptr i32 %shl0 to ptr addrspace(3)
- %ptr1 = inttoptr i32 %shl1 to ptr addrspace(3)
- %a1 = atomicrmw fadd ptr addrspace(3) %ptr0, float 4.2e+1 syncscope("one-as") seq_cst
- %a2 = atomicrmw fadd ptr addrspace(3) %ptr1, float 4.2e+1 syncscope("one-as") seq_cst
- %a3 = atomicrmw fadd ptr addrspace(3) %ptrf, float %a1 syncscope("one-as") seq_cst
- store float %a3, ptr addrspace(1) %out
- ret void
-}
-
-define double @lds_atomic_fadd_ret_f64(ptr addrspace(3) %ptr) nounwind {
-; VI-LABEL: lds_atomic_fadd_ret_f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: s_mov_b32 m0, -1
-; VI-NEXT: v_mov_b32_e32 v2, v0
-; VI-NEXT: ds_read_b64 v[0:1], v0
-; VI-NEXT: s_mov_b64 s[4:5], 0
-; VI-NEXT: .LBB4_1: ; %atomicrmw.start
-; VI-NEXT: ; =>This Inner Loop Header: Depth=1
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v4, v1
-; VI-NEXT: v_mov_b32_e32 v3, v0
-; VI-NEXT: v_add_f64 v[0:1], v[3:4], 4.0
-; VI-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
-; VI-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; VI-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB4_1
-; VI-NEXT: ; %bb.2: ; %atomicrmw.end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: lds_atomic_fadd_ret_f64:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v2, v0
-; GFX908-NEXT: ds_read_b64 v[0:1], v0
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB4_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v4, v1
-; GFX908-NEXT: v_mov_b32_e32 v3, v0
-; GFX908-NEXT: v_add_f64 v[0:1], v[3:4], 4.0
-; GFX908-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB4_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: lds_atomic_fadd_ret_f64:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v2, 0
-; GFX90A-NEXT: v_mov_b32_e32 v3, 0x40100000
-; GFX90A-NEXT: ds_add_rtn_f64 v[0:1], v0, v[2:3]
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: lds_atomic_fadd_ret_f64:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b64_e32 v[2:3], 4.0
-; GFX940-NEXT: ds_add_rtn_f64 v[0:1], v0, v[2:3]
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: lds_atomic_fadd_ret_f64:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v2, v0
-; GFX12-NEXT: ds_load_b64 v[0:1], v0
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB4_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT: v_add_f64_e32 v[0:1], 4.0, v[3:4]
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4]
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SE
-; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB4_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX7-LABEL: lds_atomic_fadd_ret_f64:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: s_mov_b32 m0, -1
-; GFX7-NEXT: v_mov_b32_e32 v2, v0
-; GFX7-NEXT: ds_read_b64 v[0:1], v0
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: .LBB4_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_mov_b32_e32 v4, v1
-; GFX7-NEXT: v_mov_b32_e32 v3, v0
-; GFX7-NEXT: v_add_f64 v[0:1], v[3:4], 4.0
-; GFX7-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB4_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: lds_atomic_fadd_ret_f64:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: s_mov_b32 m0, -1
-; GFX8-NEXT: v_mov_b32_e32 v2, v0
-; GFX8-NEXT: ds_read_b64 v[0:1], v0
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB4_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_mov_b32_e32 v4, v1
-; GFX8-NEXT: v_mov_b32_e32 v3, v0
-; GFX8-NEXT: v_add_f64 v[0:1], v[3:4], 4.0
-; GFX8-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB4_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr addrspace(3) %ptr, double 4.0 seq_cst
- ret double %result
-}
-
-define void @lds_atomic_fadd_noret_f64(ptr addrspace(3) %ptr) nounwind {
-; VI-LABEL: lds_atomic_fadd_noret_f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: s_mov_b32 m0, -1
-; VI-NEXT: ds_read_b64 v[1:2], v0
-; VI-NEXT: s_mov_b64 s[4:5], 0
-; VI-NEXT: .LBB5_1: ; %atomicrmw.start
-; VI-NEXT: ; =>This Inner Loop Header: Depth=1
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_add_f64 v[3:4], v[1:2], 4.0
-; VI-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
-; VI-NEXT: v_mov_b32_e32 v1, v3
-; VI-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v2, v4
-; VI-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB5_1
-; VI-NEXT: ; %bb.2: ; %atomicrmw.end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: lds_atomic_fadd_noret_f64:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: ds_read_b64 v[1:2], v0
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB5_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_add_f64 v[3:4], v[1:2], 4.0
-; GFX908-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
-; GFX908-NEXT: v_mov_b32_e32 v1, v3
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v2, v4
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB5_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: lds_atomic_fadd_noret_f64:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v2, 0
-; GFX90A-NEXT: v_mov_b32_e32 v3, 0x40100000
-; GFX90A-NEXT: ds_add_f64 v0, v[2:3]
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: lds_atomic_fadd_noret_f64:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b64_e32 v[2:3], 4.0
-; GFX940-NEXT: ds_add_f64 v0, v[2:3]
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: lds_atomic_fadd_noret_f64:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: ds_load_b64 v[1:2], v0
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB5_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: v_add_f64_e32 v[3:4], 4.0, v[1:2]
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2]
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SE
-; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
-; GFX12-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB5_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX7-LABEL: lds_atomic_fadd_noret_f64:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: s_mov_b32 m0, -1
-; GFX7-NEXT: ds_read_b64 v[1:2], v0
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: .LBB5_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_add_f64 v[3:4], v[1:2], 4.0
-; GFX7-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
-; GFX7-NEXT: v_mov_b32_e32 v1, v3
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: v_mov_b32_e32 v2, v4
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB5_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: lds_atomic_fadd_noret_f64:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: s_mov_b32 m0, -1
-; GFX8-NEXT: ds_read_b64 v[1:2], v0
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB5_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_add_f64 v[3:4], v[1:2], 4.0
-; GFX8-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
-; GFX8-NEXT: v_mov_b32_e32 v1, v3
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v2, v4
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB5_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr addrspace(3) %ptr, double 4.0 seq_cst
- ret void
-}
-
-define float @lds_atomic_fsub_ret_f32(ptr addrspace(3) %ptr, float %val) nounwind {
-; VI-LABEL: lds_atomic_fsub_ret_f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: s_mov_b32 m0, -1
-; VI-NEXT: ds_read_b32 v2, v0
-; VI-NEXT: s_mov_b64 s[4:5], 0
-; VI-NEXT: .LBB6_1: ; %atomicrmw.start
-; VI-NEXT: ; =>This Inner Loop Header: Depth=1
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v3, v2
-; VI-NEXT: v_sub_f32_e32 v2, v3, v1
-; VI-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; VI-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; VI-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB6_1
-; VI-NEXT: ; %bb.2: ; %atomicrmw.end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v0, v2
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: lds_atomic_fsub_ret_f32:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: ds_read_b32 v2, v0
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB6_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v3, v2
-; GFX908-NEXT: v_sub_f32_e32 v2, v3, v1
-; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB6_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v0, v2
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: lds_atomic_fsub_ret_f32:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: ds_read_b32 v2, v0
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB6_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v3, v2
-; GFX90A-NEXT: v_sub_f32_e32 v2, v3, v1
-; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB6_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v2
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: lds_atomic_fsub_ret_f32:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: ds_read_b32 v2, v0
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB6_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: v_sub_f32_e32 v2, v3, v1
-; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB6_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v0, v2
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: lds_atomic_fsub_ret_f32:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: ds_load_b32 v2, v0
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB6_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v3, v2
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT: v_sub_f32_e32 v2, v3, v1
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SE
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB6_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: v_mov_b32_e32 v0, v2
-; GFX12-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX7-LABEL: lds_atomic_fsub_ret_f32:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: s_mov_b32 m0, -1
-; GFX7-NEXT: ds_read_b32 v2, v0
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: .LBB6_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_mov_b32_e32 v3, v2
-; GFX7-NEXT: v_sub_f32_e32 v2, v3, v1
-; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB6_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: v_mov_b32_e32 v0, v2
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: lds_atomic_fsub_ret_f32:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: s_mov_b32 m0, -1
-; GFX8-NEXT: ds_read_b32 v2, v0
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB6_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_mov_b32_e32 v3, v2
-; GFX8-NEXT: v_sub_f32_e32 v2, v3, v1
-; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB6_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v0, v2
-; GFX8-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fsub ptr addrspace(3) %ptr, float %val seq_cst
- ret float %result
-}
-
-define void @lds_atomic_fsub_noret_f32(ptr addrspace(3) %ptr, float %val) nounwind {
-; VI-LABEL: lds_atomic_fsub_noret_f32:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: s_mov_b32 m0, -1
-; VI-NEXT: ds_read_b32 v2, v0
-; VI-NEXT: s_mov_b64 s[4:5], 0
-; VI-NEXT: .LBB7_1: ; %atomicrmw.start
-; VI-NEXT: ; =>This Inner Loop Header: Depth=1
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_sub_f32_e32 v3, v2, v1
-; VI-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
-; VI-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v2, v3
-; VI-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB7_1
-; VI-NEXT: ; %bb.2: ; %atomicrmw.end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: lds_atomic_fsub_noret_f32:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: ds_read_b32 v2, v0
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB7_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_sub_f32_e32 v3, v2, v1
-; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v2, v3
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB7_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: lds_atomic_fsub_noret_f32:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: ds_read_b32 v2, v0
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB7_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_sub_f32_e32 v3, v2, v1
-; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB7_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: lds_atomic_fsub_noret_f32:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: ds_read_b32 v2, v0
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB7_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_sub_f32_e32 v3, v2, v1
-; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB7_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: lds_atomic_fsub_noret_f32:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: ds_load_b32 v2, v0
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB7_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: v_sub_f32_e32 v3, v2, v1
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SE
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
-; GFX12-NEXT: v_mov_b32_e32 v2, v3
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB7_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX7-LABEL: lds_atomic_fsub_noret_f32:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: s_mov_b32 m0, -1
-; GFX7-NEXT: ds_read_b32 v2, v0
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: .LBB7_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_sub_f32_e32 v3, v2, v1
-; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: v_mov_b32_e32 v2, v3
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB7_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: lds_atomic_fsub_noret_f32:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: s_mov_b32 m0, -1
-; GFX8-NEXT: ds_read_b32 v2, v0
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB7_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_sub_f32_e32 v3, v2, v1
-; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v2, v3
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB7_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fsub ptr addrspace(3) %ptr, float %val seq_cst
- ret void
-}
-
-define double @lds_atomic_fsub_ret_f64(ptr addrspace(3) %ptr, double %val) nounwind {
-; VI-LABEL: lds_atomic_fsub_ret_f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: s_mov_b32 m0, -1
-; VI-NEXT: ds_read_b64 v[3:4], v0
-; VI-NEXT: s_mov_b64 s[4:5], 0
-; VI-NEXT: .LBB8_1: ; %atomicrmw.start
-; VI-NEXT: ; =>This Inner Loop Header: Depth=1
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v6, v4
-; VI-NEXT: v_mov_b32_e32 v5, v3
-; VI-NEXT: v_add_f64 v[3:4], v[5:6], -v[1:2]
-; VI-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[5:6], v[3:4]
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
-; VI-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; VI-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB8_1
-; VI-NEXT: ; %bb.2: ; %atomicrmw.end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v0, v3
-; VI-NEXT: v_mov_b32_e32 v1, v4
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: lds_atomic_fsub_ret_f64:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: ds_read_b64 v[3:4], v0
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB8_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v6, v4
-; GFX908-NEXT: v_mov_b32_e32 v5, v3
-; GFX908-NEXT: v_add_f64 v[3:4], v[5:6], -v[1:2]
-; GFX908-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[5:6], v[3:4]
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB8_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v0, v3
-; GFX908-NEXT: v_mov_b32_e32 v1, v4
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: lds_atomic_fsub_ret_f64:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
-; GFX90A-NEXT: v_mov_b32_e32 v2, v0
-; GFX90A-NEXT: v_mov_b32_e32 v4, v1
-; GFX90A-NEXT: ds_read_b64 v[0:1], v0
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB8_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[0:1], v[0:1] op_sel:[0,1]
-; GFX90A-NEXT: v_add_f64 v[0:1], v[6:7], -v[4:5]
-; GFX90A-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[6:7], v[0:1]
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[6:7]
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB8_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: lds_atomic_fsub_ret_f64:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
-; GFX940-NEXT: v_mov_b32_e32 v2, v0
-; GFX940-NEXT: v_mov_b32_e32 v4, v1
-; GFX940-NEXT: ds_read_b64 v[0:1], v0
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB8_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[0:1]
-; GFX940-NEXT: v_add_f64 v[0:1], v[6:7], -v[4:5]
-; GFX940-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[6:7], v[0:1]
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[6:7]
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB8_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: lds_atomic_fsub_ret_f64:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: ds_load_b64 v[3:4], v0
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB8_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: v_dual_mov_b32 v6, v4 :: v_dual_mov_b32 v5, v3
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT: v_add_f64_e64 v[3:4], v[5:6], -v[1:2]
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[5:6]
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SE
-; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[5:6]
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB8_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: v_dual_mov_b32 v0, v3 :: v_dual_mov_b32 v1, v4
-; GFX12-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX7-LABEL: lds_atomic_fsub_ret_f64:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: s_mov_b32 m0, -1
-; GFX7-NEXT: ds_read_b64 v[3:4], v0
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: .LBB8_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_mov_b32_e32 v6, v4
-; GFX7-NEXT: v_mov_b32_e32 v5, v3
-; GFX7-NEXT: v_add_f64 v[3:4], v[5:6], -v[1:2]
-; GFX7-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[5:6], v[3:4]
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB8_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: v_mov_b32_e32 v0, v3
-; GFX7-NEXT: v_mov_b32_e32 v1, v4
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: lds_atomic_fsub_ret_f64:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: s_mov_b32 m0, -1
-; GFX8-NEXT: ds_read_b64 v[3:4], v0
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB8_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_mov_b32_e32 v6, v4
-; GFX8-NEXT: v_mov_b32_e32 v5, v3
-; GFX8-NEXT: v_add_f64 v[3:4], v[5:6], -v[1:2]
-; GFX8-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[5:6], v[3:4]
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB8_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v0, v3
-; GFX8-NEXT: v_mov_b32_e32 v1, v4
-; GFX8-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fsub ptr addrspace(3) %ptr, double %val seq_cst
- ret double %result
-}
-
-define void @lds_atomic_fsub_noret_f64(ptr addrspace(3) %ptr, double %val) nounwind {
-; VI-LABEL: lds_atomic_fsub_noret_f64:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: s_mov_b32 m0, -1
-; VI-NEXT: ds_read_b64 v[3:4], v0
-; VI-NEXT: s_mov_b64 s[4:5], 0
-; VI-NEXT: .LBB9_1: ; %atomicrmw.start
-; VI-NEXT: ; =>This Inner Loop Header: Depth=1
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_add_f64 v[5:6], v[3:4], -v[1:2]
-; VI-NEXT: ds_cmpst_rtn_b64 v[5:6], v0, v[3:4], v[5:6]
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_cmp_eq_u64_e32 vcc, v[5:6], v[3:4]
-; VI-NEXT: v_mov_b32_e32 v3, v5
-; VI-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v4, v6
-; VI-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB9_1
-; VI-NEXT: ; %bb.2: ; %atomicrmw.end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: lds_atomic_fsub_noret_f64:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: ds_read_b64 v[3:4], v0
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB9_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_add_f64 v[5:6], v[3:4], -v[1:2]
-; GFX908-NEXT: ds_cmpst_rtn_b64 v[5:6], v0, v[3:4], v[5:6]
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[5:6], v[3:4]
-; GFX908-NEXT: v_mov_b32_e32 v3, v5
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v4, v6
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB9_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: lds_atomic_fsub_noret_f64:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: ds_read_b64 v[4:5], v0
-; GFX90A-NEXT: v_mov_b32_e32 v3, v2
-; GFX90A-NEXT: v_mov_b32_e32 v2, v1
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB9_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_add_f64 v[6:7], v[4:5], -v[2:3]
-; GFX90A-NEXT: ds_cmpst_rtn_b64 v[6:7], v0, v[4:5], v[6:7]
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[4:5]
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[6:7] op_sel:[0,1]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB9_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: lds_atomic_fsub_noret_f64:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: ds_read_b64 v[4:5], v0
-; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: v_mov_b32_e32 v2, v1
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB9_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_add_f64 v[6:7], v[4:5], -v[2:3]
-; GFX940-NEXT: ds_cmpst_rtn_b64 v[6:7], v0, v[4:5], v[6:7]
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[4:5]
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[6:7]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB9_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: lds_atomic_fsub_noret_f64:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: ds_load_b64 v[3:4], v0
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB9_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: v_add_f64_e64 v[5:6], v[3:4], -v[1:2]
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: ds_cmpstore_rtn_b64 v[5:6], v0, v[5:6], v[3:4]
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SE
-; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[5:6], v[3:4]
-; GFX12-NEXT: v_dual_mov_b32 v3, v5 :: v_dual_mov_b32 v4, v6
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB9_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX7-LABEL: lds_atomic_fsub_noret_f64:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: s_mov_b32 m0, -1
-; GFX7-NEXT: ds_read_b64 v[3:4], v0
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: .LBB9_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_add_f64 v[5:6], v[3:4], -v[1:2]
-; GFX7-NEXT: ds_cmpst_rtn_b64 v[5:6], v0, v[3:4], v[5:6]
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[5:6], v[3:4]
-; GFX7-NEXT: v_mov_b32_e32 v3, v5
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: v_mov_b32_e32 v4, v6
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB9_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: lds_atomic_fsub_noret_f64:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: s_mov_b32 m0, -1
-; GFX8-NEXT: ds_read_b64 v[3:4], v0
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB9_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_add_f64 v[5:6], v[3:4], -v[1:2]
-; GFX8-NEXT: ds_cmpst_rtn_b64 v[5:6], v0, v[3:4], v[5:6]
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[5:6], v[3:4]
-; GFX8-NEXT: v_mov_b32_e32 v3, v5
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v4, v6
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB9_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fsub ptr addrspace(3) %ptr, double %val seq_cst
- ret void
-}
-
-define bfloat @lds_atomic_fadd_ret_bf16(ptr addrspace(3) %ptr) nounwind {
-; VI-LABEL: lds_atomic_fadd_ret_bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_and_b32_e32 v1, -4, v0
-; VI-NEXT: s_mov_b32 m0, -1
-; VI-NEXT: ds_read_b32 v3, v1
-; VI-NEXT: v_lshlrev_b32_e32 v2, 3, v0
-; VI-NEXT: s_mov_b32 s4, 0xffff
-; VI-NEXT: v_and_b32_e32 v0, 24, v2
-; VI-NEXT: v_lshlrev_b32_e64 v2, v2, s4
-; VI-NEXT: v_not_b32_e32 v2, v2
-; VI-NEXT: s_mov_b64 s[4:5], 0
-; VI-NEXT: .LBB10_1: ; %atomicrmw.start
-; VI-NEXT: ; =>This Inner Loop Header: Depth=1
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v4, v3
-; VI-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_f32_e32 v3, 4.0, v3
-; VI-NEXT: v_bfe_u32 v6, v3, 16, 1
-; VI-NEXT: v_add_u32_e32 v6, vcc, v6, v3
-; VI-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
-; VI-NEXT: v_or_b32_e32 v7, 0x400000, v3
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; VI-NEXT: v_cndmask_b32_e32 v3, v6, v7, vcc
-; VI-NEXT: v_and_b32_e32 v5, v4, v2
-; VI-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; VI-NEXT: v_or_b32_e32 v3, v5, v3
-; VI-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; VI-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; VI-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB10_1
-; VI-NEXT: ; %bb.2: ; %atomicrmw.end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_lshrrev_b32_e32 v0, v0, v3
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: lds_atomic_fadd_ret_bf16:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
-; GFX908-NEXT: ds_read_b32 v3, v1
-; GFX908-NEXT: v_lshlrev_b32_e32 v2, 3, v0
-; GFX908-NEXT: s_mov_b32 s4, 0xffff
-; GFX908-NEXT: v_and_b32_e32 v0, 24, v2
-; GFX908-NEXT: v_lshlrev_b32_e64 v2, v2, s4
-; GFX908-NEXT: v_not_b32_e32 v2, v2
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: s_movk_i32 s6, 0x7fff
-; GFX908-NEXT: .LBB10_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v4, v3
-; GFX908-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX908-NEXT: v_add_f32_e32 v3, 4.0, v3
-; GFX908-NEXT: v_bfe_u32 v5, v3, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v3
-; GFX908-NEXT: v_add3_u32 v5, v5, v3, s6
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX908-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
-; GFX908-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; GFX908-NEXT: v_and_or_b32 v3, v4, v2, v3
-; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB10_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: v_lshrrev_b32_e32 v0, v0, v3
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: lds_atomic_fadd_ret_bf16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
-; GFX90A-NEXT: ds_read_b32 v3, v1
-; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 3, v0
-; GFX90A-NEXT: s_mov_b32 s4, 0xffff
-; GFX90A-NEXT: v_and_b32_e32 v0, 24, v2
-; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v2, s4
-; GFX90A-NEXT: v_not_b32_e32 v2, v2
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
-; GFX90A-NEXT: .LBB10_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX90A-NEXT: v_add_f32_e32 v3, 4.0, v3
-; GFX90A-NEXT: v_bfe_u32 v5, v3, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3
-; GFX90A-NEXT: v_add3_u32 v5, v5, v3, s6
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX90A-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
-; GFX90A-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; GFX90A-NEXT: v_and_or_b32 v3, v4, v2, v3
-; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB10_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v3
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: lds_atomic_fadd_ret_bf16:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
-; GFX940-NEXT: ds_read_b32 v3, v1
-; GFX940-NEXT: v_lshlrev_b32_e32 v2, 3, v0
-; GFX940-NEXT: s_mov_b32 s0, 0xffff
-; GFX940-NEXT: v_and_b32_e32 v0, 24, v2
-; GFX940-NEXT: v_lshlrev_b32_e64 v2, v2, s0
-; GFX940-NEXT: v_not_b32_e32 v2, v2
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: s_movk_i32 s2, 0x7fff
-; GFX940-NEXT: .LBB10_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_add_f32_e32 v3, 4.0, v3
-; GFX940-NEXT: v_bfe_u32 v5, v3, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v3
-; GFX940-NEXT: v_add3_u32 v5, v5, v3, s2
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX940-NEXT: s_nop 1
-; GFX940-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
-; GFX940-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; GFX940-NEXT: v_and_or_b32 v3, v4, v2, v3
-; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB10_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: v_lshrrev_b32_e32 v0, v0, v3
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: lds_atomic_fadd_ret_bf16:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: v_and_b32_e32 v1, -4, v0
-; GFX12-NEXT: v_lshlrev_b32_e32 v0, 3, v0
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: ds_load_b32 v3, v1
-; GFX12-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff
-; GFX12-NEXT: v_and_b32_e32 v0, 24, v0
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX12-NEXT: v_not_b32_e32 v2, v2
-; GFX12-NEXT: .LBB10_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT: v_lshrrev_b32_e32 v3, v0, v4
-; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT: v_add_f32_e32 v3, 4.0, v3
-; GFX12-NEXT: v_bfe_u32 v5, v3, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v3
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
-; GFX12-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GFX12-NEXT: v_lshlrev_b32_e32 v3, v0, v3
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT: v_and_or_b32 v3, v4, v2, v3
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SE
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB10_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: v_lshrrev_b32_e32 v0, v0, v3
-; GFX12-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX7-LABEL: lds_atomic_fadd_ret_bf16:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_and_b32_e32 v1, -4, v0
-; GFX7-NEXT: s_mov_b32 m0, -1
-; GFX7-NEXT: ds_read_b32 v3, v1
-; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v0
-; GFX7-NEXT: v_and_b32_e32 v0, 24, v2
-; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
-; GFX7-NEXT: v_not_b32_e32 v2, v2
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: .LBB10_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_mov_b32_e32 v4, v3
-; GFX7-NEXT: v_lshrrev_b32_e32 v3, v0, v4
-; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX7-NEXT: v_add_f32_e32 v3, 4.0, v3
-; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GFX7-NEXT: v_and_b32_e32 v5, v4, v2
-; GFX7-NEXT: v_lshlrev_b32_e32 v3, v0, v3
-; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
-; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB10_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: v_lshrrev_b32_e32 v0, v0, v3
-; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: lds_atomic_fadd_ret_bf16:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
-; GFX8-NEXT: s_mov_b32 m0, -1
-; GFX8-NEXT: ds_read_b32 v3, v1
-; GFX8-NEXT: v_lshlrev_b32_e32 v2, 3, v0
-; GFX8-NEXT: v_and_b32_e32 v0, 24, v2
-; GFX8-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
-; GFX8-NEXT: v_not_b32_e32 v2, v2
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB10_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_mov_b32_e32 v4, v3
-; GFX8-NEXT: v_lshrrev_b32_e32 v3, v0, v4
-; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX8-NEXT: v_add_f32_e32 v3, 4.0, v3
-; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GFX8-NEXT: v_and_b32_e32 v5, v4, v2
-; GFX8-NEXT: v_lshlrev_b32_e32 v3, v0, v3
-; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
-; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB10_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: v_lshrrev_b32_e32 v0, v0, v3
-; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GFX8-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr addrspace(3) %ptr, bfloat 4.0 seq_cst
- ret bfloat %result
-}
-
-define void @lds_atomic_fadd_noret_bf16(ptr addrspace(3) %ptr) nounwind {
-; VI-LABEL: lds_atomic_fadd_noret_bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_and_b32_e32 v1, -4, v0
-; VI-NEXT: s_mov_b32 m0, -1
-; VI-NEXT: ds_read_b32 v3, v1
-; VI-NEXT: v_lshlrev_b32_e32 v2, 3, v0
-; VI-NEXT: s_mov_b32 s4, 0xffff
-; VI-NEXT: v_and_b32_e32 v0, 24, v2
-; VI-NEXT: v_lshlrev_b32_e64 v2, v2, s4
-; VI-NEXT: v_not_b32_e32 v2, v2
-; VI-NEXT: s_mov_b64 s[4:5], 0
-; VI-NEXT: .LBB11_1: ; %atomicrmw.start
-; VI-NEXT: ; =>This Inner Loop Header: Depth=1
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; VI-NEXT: v_add_f32_e32 v4, 4.0, v4
-; VI-NEXT: v_bfe_u32 v6, v4, 16, 1
-; VI-NEXT: v_add_u32_e32 v6, vcc, v6, v4
-; VI-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
-; VI-NEXT: v_or_b32_e32 v7, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v4, v6, v7, vcc
-; VI-NEXT: v_and_b32_e32 v5, v3, v2
-; VI-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; VI-NEXT: v_or_b32_e32 v4, v5, v4
-; VI-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
-; VI-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v3, v4
-; VI-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB11_1
-; VI-NEXT: ; %bb.2: ; %atomicrmw.end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: lds_atomic_fadd_noret_bf16:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_and_b32_e32 v1, -4, v0
-; GFX908-NEXT: ds_read_b32 v3, v1
-; GFX908-NEXT: v_lshlrev_b32_e32 v2, 3, v0
-; GFX908-NEXT: s_mov_b32 s4, 0xffff
-; GFX908-NEXT: v_and_b32_e32 v0, 24, v2
-; GFX908-NEXT: v_lshlrev_b32_e64 v2, v2, s4
-; GFX908-NEXT: v_not_b32_e32 v2, v2
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: s_movk_i32 s6, 0x7fff
-; GFX908-NEXT: .LBB11_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX908-NEXT: v_add_f32_e32 v4, 4.0, v4
-; GFX908-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX908-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX908-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX908-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4
-; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v3, v4
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB11_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: lds_atomic_fadd_noret_bf16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0
-; GFX90A-NEXT: ds_read_b32 v3, v1
-; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 3, v0
-; GFX90A-NEXT: s_mov_b32 s4, 0xffff
-; GFX90A-NEXT: v_and_b32_e32 v0, 24, v2
-; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v2, s4
-; GFX90A-NEXT: v_not_b32_e32 v2, v2
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
-; GFX90A-NEXT: .LBB11_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX90A-NEXT: v_add_f32_e32 v4, 4.0, v4
-; GFX90A-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX90A-NEXT: v_add3_u32 v5, v5, v4, s6
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX90A-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4
-; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB11_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: lds_atomic_fadd_noret_bf16:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_and_b32_e32 v1, -4, v0
-; GFX940-NEXT: ds_read_b32 v3, v1
-; GFX940-NEXT: v_lshlrev_b32_e32 v2, 3, v0
-; GFX940-NEXT: s_mov_b32 s0, 0xffff
-; GFX940-NEXT: v_and_b32_e32 v0, 24, v2
-; GFX940-NEXT: v_lshlrev_b32_e64 v2, v2, s0
-; GFX940-NEXT: v_not_b32_e32 v2, v2
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: s_movk_i32 s2, 0x7fff
-; GFX940-NEXT: .LBB11_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_add_f32_e32 v4, 4.0, v4
-; GFX940-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX940-NEXT: v_add3_u32 v5, v5, v4, s2
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
-; GFX940-NEXT: s_nop 1
-; GFX940-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
-; GFX940-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; GFX940-NEXT: v_and_or_b32 v4, v3, v2, v4
-; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB11_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: lds_atomic_fadd_noret_bf16:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: v_and_b32_e32 v1, -4, v0
-; GFX12-NEXT: v_lshlrev_b32_e32 v0, 3, v0
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: ds_load_b32 v2, v1
-; GFX12-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff
-; GFX12-NEXT: v_and_b32_e32 v0, 24, v0
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX12-NEXT: v_not_b32_e32 v3, v3
-; GFX12-NEXT: .LBB11_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT: v_lshrrev_b32_e32 v4, v0, v2
-; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT: v_add_f32_e32 v4, 4.0, v4
-; GFX12-NEXT: v_bfe_u32 v5, v4, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v4
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
-; GFX12-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GFX12-NEXT: v_lshlrev_b32_e32 v4, v0, v4
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT: v_and_or_b32 v4, v2, v3, v4
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SE
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
-; GFX12-NEXT: v_mov_b32_e32 v2, v4
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB11_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX7-LABEL: lds_atomic_fadd_noret_bf16:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_and_b32_e32 v1, -4, v0
-; GFX7-NEXT: s_mov_b32 m0, -1
-; GFX7-NEXT: ds_read_b32 v3, v1
-; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v0
-; GFX7-NEXT: v_and_b32_e32 v0, 24, v2
-; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
-; GFX7-NEXT: v_not_b32_e32 v2, v2
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: .LBB11_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_lshrrev_b32_e32 v4, v0, v3
-; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX7-NEXT: v_add_f32_e32 v4, 4.0, v4
-; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GFX7-NEXT: v_and_b32_e32 v5, v3, v2
-; GFX7-NEXT: v_lshlrev_b32_e32 v4, v0, v4
-; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
-; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: v_mov_b32_e32 v3, v4
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB11_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: lds_atomic_fadd_noret_bf16:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_and_b32_e32 v1, -4, v0
-; GFX8-NEXT: s_mov_b32 m0, -1
-; GFX8-NEXT: ds_read_b32 v3, v1
-; GFX8-NEXT: v_lshlrev_b32_e32 v2, 3, v0
-; GFX8-NEXT: v_and_b32_e32 v0, 24, v2
-; GFX8-NEXT: v_lshl_b32_e32 v2, 0xffff, v2
-; GFX8-NEXT: v_not_b32_e32 v2, v2
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB11_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_lshrrev_b32_e32 v4, v0, v3
-; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX8-NEXT: v_add_f32_e32 v4, 4.0, v4
-; GFX8-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GFX8-NEXT: v_and_b32_e32 v5, v3, v2
-; GFX8-NEXT: v_lshlrev_b32_e32 v4, v0, v4
-; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
-; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v3, v4
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB11_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr addrspace(3) %ptr, bfloat 4.0 seq_cst
- ret void
-}
-
-define float @lds_atomic_fadd_ret_f32__amdgpu_ignore_denormal_mode(ptr addrspace(3) %ptr) {
-; VI-LABEL: lds_atomic_fadd_ret_f32__amdgpu_ignore_denormal_mode:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v1, 4.0
-; VI-NEXT: s_mov_b32 m0, -1
-; VI-NEXT: ds_add_rtn_f32 v0, v0, v1
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: lds_atomic_fadd_ret_f32__amdgpu_ignore_denormal_mode:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v1, 4.0
-; GFX9-NEXT: ds_add_rtn_f32 v0, v0, v1
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: lds_atomic_fadd_ret_f32__amdgpu_ignore_denormal_mode:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v1, 4.0
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: ds_add_rtn_f32 v0, v0, v1
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SE
-; GFX12-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX7-LABEL: lds_atomic_fadd_ret_f32__amdgpu_ignore_denormal_mode:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: s_mov_b32 m0, -1
-; GFX7-NEXT: ds_read_b32 v1, v0
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: .LBB12_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_mov_b32_e32 v2, v1
-; GFX7-NEXT: v_add_f32_e32 v1, 4.0, v2
-; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB12_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: v_mov_b32_e32 v0, v1
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: lds_atomic_fadd_ret_f32__amdgpu_ignore_denormal_mode:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: s_mov_b32 m0, -1
-; GFX8-NEXT: ds_read_b32 v1, v0
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB12_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_mov_b32_e32 v2, v1
-; GFX8-NEXT: v_add_f32_e32 v1, 4.0, v2
-; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB12_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v0, v1
-; GFX8-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr addrspace(3) %ptr, float 4.0 seq_cst, !amdgpu.ignore.denormal.mode !0
- ret float %result
-}
-
-define void @lds_atomic_fadd_noret_f32__amdgpu_ignore_denormal_mode(ptr addrspace(3) %ptr) {
-; VI-LABEL: lds_atomic_fadd_noret_f32__amdgpu_ignore_denormal_mode:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v1, 4.0
-; VI-NEXT: s_mov_b32 m0, -1
-; VI-NEXT: ds_add_f32 v0, v1
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX9-LABEL: lds_atomic_fadd_noret_f32__amdgpu_ignore_denormal_mode:
-; GFX9: ; %bb.0:
-; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v1, 4.0
-; GFX9-NEXT: ds_add_f32 v0, v1
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: lds_atomic_fadd_noret_f32__amdgpu_ignore_denormal_mode:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v1, 4.0
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: ds_add_f32 v0, v1
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SE
-; GFX12-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX7-LABEL: lds_atomic_fadd_noret_f32__amdgpu_ignore_denormal_mode:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: s_mov_b32 m0, -1
-; GFX7-NEXT: ds_read_b32 v1, v0
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: .LBB13_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_add_f32_e32 v2, 4.0, v1
-; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: v_mov_b32_e32 v1, v2
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB13_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: lds_atomic_fadd_noret_f32__amdgpu_ignore_denormal_mode:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: s_mov_b32 m0, -1
-; GFX8-NEXT: ds_read_b32 v1, v0
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB13_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_add_f32_e32 v2, 4.0, v1
-; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v1, v2
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB13_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr addrspace(3) %ptr, float 4.0 seq_cst, !amdgpu.ignore.denormal.mode !0
- ret void
-}
-
-define <2 x half> @lds_atomic_fadd_ret_v2f16(ptr addrspace(3) %ptr, <2 x half> %val) {
-; VI-LABEL: lds_atomic_fadd_ret_v2f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: s_mov_b32 m0, -1
-; VI-NEXT: ds_read_b32 v2, v0
-; VI-NEXT: s_mov_b64 s[4:5], 0
-; VI-NEXT: .LBB14_1: ; %atomicrmw.start
-; VI-NEXT: ; =>This Inner Loop Header: Depth=1
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v3, v2
-; VI-NEXT: v_add_f16_sdwa v2, v3, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
-; VI-NEXT: v_add_f16_e32 v4, v3, v1
-; VI-NEXT: v_or_b32_e32 v2, v4, v2
-; VI-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; VI-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; VI-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB14_1
-; VI-NEXT: ; %bb.2: ; %atomicrmw.end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v0, v2
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: lds_atomic_fadd_ret_v2f16:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: ds_read_b32 v2, v0
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB14_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v3, v2
-; GFX908-NEXT: v_pk_add_f16 v2, v3, v1
-; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB14_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v0, v2
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: lds_atomic_fadd_ret_v2f16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: ds_read_b32 v2, v0
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB14_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v3, v2
-; GFX90A-NEXT: v_pk_add_f16 v2, v3, v1
-; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB14_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v2
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: lds_atomic_fadd_ret_v2f16:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: ds_read_b32 v2, v0
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB14_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: v_pk_add_f16 v2, v3, v1
-; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB14_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v0, v2
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: lds_atomic_fadd_ret_v2f16:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: ds_load_b32 v2, v0
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB14_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v3, v2
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT: v_pk_add_f16 v2, v3, v1
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SE
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB14_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: v_mov_b32_e32 v0, v2
-; GFX12-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX7-LABEL: lds_atomic_fadd_ret_v2f16:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: s_mov_b32 m0, -1
-; GFX7-NEXT: ds_read_b32 v3, v0
-; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
-; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v1
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
-; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
-; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
-; GFX7-NEXT: .LBB14_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v1
-; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v2
-; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX7-NEXT: v_or_b32_e32 v7, v2, v1
-; GFX7-NEXT: v_add_f32_e32 v5, v5, v3
-; GFX7-NEXT: v_add_f32_e32 v6, v6, v4
-; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v5
-; GFX7-NEXT: v_or_b32_e32 v1, v6, v1
-; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB14_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: v_mov_b32_e32 v0, v2
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: lds_atomic_fadd_ret_v2f16:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: s_mov_b32 m0, -1
-; GFX8-NEXT: ds_read_b32 v3, v0
-; GFX8-NEXT: v_cvt_f16_f32_e32 v4, v2
-; GFX8-NEXT: v_cvt_f16_f32_e32 v5, v1
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v3
-; GFX8-NEXT: v_cvt_f32_f16_e32 v1, v2
-; GFX8-NEXT: v_cvt_f32_f16_e32 v2, v3
-; GFX8-NEXT: v_cvt_f32_f16_e32 v3, v4
-; GFX8-NEXT: v_cvt_f32_f16_e32 v4, v5
-; GFX8-NEXT: .LBB14_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GFX8-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GFX8-NEXT: v_cvt_f32_f16_e32 v5, v1
-; GFX8-NEXT: v_cvt_f32_f16_e32 v6, v2
-; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX8-NEXT: v_or_b32_e32 v7, v2, v1
-; GFX8-NEXT: v_add_f32_e32 v5, v5, v3
-; GFX8-NEXT: v_add_f32_e32 v6, v6, v4
-; GFX8-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GFX8-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v5
-; GFX8-NEXT: v_or_b32_e32 v1, v6, v1
-; GFX8-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v5
-; GFX8-NEXT: v_cvt_f32_f16_e32 v2, v5
-; GFX8-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB14_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v0, v2
-; GFX8-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr addrspace(3) %ptr, <2 x half> %val seq_cst
- ret <2 x half> %result
-}
-
-define void @lds_atomic_fadd_noret_v2f16(ptr addrspace(3) %ptr, <2 x half> %val) {
-; VI-LABEL: lds_atomic_fadd_noret_v2f16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: s_mov_b32 m0, -1
-; VI-NEXT: ds_read_b32 v2, v0
-; VI-NEXT: s_mov_b64 s[4:5], 0
-; VI-NEXT: .LBB15_1: ; %atomicrmw.start
-; VI-NEXT: ; =>This Inner Loop Header: Depth=1
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_add_f16_sdwa v3, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
-; VI-NEXT: v_add_f16_e32 v4, v2, v1
-; VI-NEXT: v_or_b32_e32 v3, v4, v3
-; VI-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
-; VI-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; VI-NEXT: v_mov_b32_e32 v2, v3
-; VI-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; VI-NEXT: s_cbranch_execnz .LBB15_1
-; VI-NEXT: ; %bb.2: ; %atomicrmw.end
-; VI-NEXT: s_or_b64 exec, exec, s[4:5]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: lds_atomic_fadd_noret_v2f16:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: ds_read_b32 v2, v0
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB15_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_pk_add_f16 v3, v2, v1
-; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v2, v3
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB15_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: lds_atomic_fadd_noret_v2f16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: ds_read_b32 v2, v0
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB15_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_pk_add_f16 v3, v2, v1
-; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB15_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: lds_atomic_fadd_noret_v2f16:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: ds_read_b32 v2, v0
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB15_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_pk_add_f16 v3, v2, v1
-; GFX940-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB15_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: lds_atomic_fadd_noret_v2f16:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: ds_load_b32 v2, v0
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB15_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: v_pk_add_f16 v3, v2, v1
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SE
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
-; GFX12-NEXT: v_mov_b32_e32 v2, v3
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB15_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX7-LABEL: lds_atomic_fadd_noret_v2f16:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: s_mov_b32 m0, -1
-; GFX7-NEXT: ds_read_b32 v3, v0
-; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v1
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
-; GFX7-NEXT: .LBB15_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v4
-; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
-; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX7-NEXT: v_or_b32_e32 v7, v3, v4
-; GFX7-NEXT: v_add_f32_e32 v5, v5, v1
-; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
-; GFX7-NEXT: v_or_b32_e32 v3, v6, v3
-; GFX7-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v3
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB15_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: lds_atomic_fadd_noret_v2f16:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: s_mov_b32 m0, -1
-; GFX8-NEXT: ds_read_b32 v3, v0
-; GFX8-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GFX8-NEXT: v_cvt_f16_f32_e32 v5, v1
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: v_cvt_f32_f16_e32 v1, v2
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_lshrrev_b32_e32 v4, 16, v3
-; GFX8-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GFX8-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GFX8-NEXT: v_cvt_f32_f16_e32 v2, v5
-; GFX8-NEXT: .LBB15_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GFX8-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GFX8-NEXT: v_cvt_f32_f16_e32 v5, v4
-; GFX8-NEXT: v_cvt_f32_f16_e32 v6, v3
-; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX8-NEXT: v_or_b32_e32 v7, v3, v4
-; GFX8-NEXT: v_add_f32_e32 v5, v5, v1
-; GFX8-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX8-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GFX8-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v5
-; GFX8-NEXT: v_or_b32_e32 v3, v6, v3
-; GFX8-NEXT: ds_cmpst_rtn_b32 v5, v0, v7, v3
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_lshrrev_b32_e32 v4, 16, v5
-; GFX8-NEXT: v_cvt_f32_f16_e32 v3, v5
-; GFX8-NEXT: v_cvt_f32_f16_e32 v4, v4
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB15_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr addrspace(3) %ptr, <2 x half> %val seq_cst
- ret void
-}
-
-define <2 x bfloat> @lds_atomic_fadd_ret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
-; VI-LABEL: lds_atomic_fadd_ret_v2bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: s_mov_b32 m0, -1
-; VI-NEXT: ds_read_b32 v2, v0
-; VI-NEXT: s_mov_b64 s[6:7], 0
-; VI-NEXT: v_lshlrev_b32_e32 v3, 16, v1
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: .LBB16_1: ; %atomicrmw.start
-; VI-NEXT: ; =>This Inner Loop Header: Depth=1
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_mov_b32_e32 v4, v2
-; VI-NEXT: v_lshlrev_b32_e32 v2, 16, v4
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
-; VI-NEXT: v_add_f32_e32 v2, v2, v3
-; VI-NEXT: v_add_f32_e32 v5, v5, v1
-; VI-NEXT: v_bfe_u32 v6, v2, 16, 1
-; VI-NEXT: v_bfe_u32 v8, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v6, vcc, v6, v2
-; VI-NEXT: v_add_u32_e32 v8, vcc, v8, v5
-; VI-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
-; VI-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
-; VI-NEXT: v_or_b32_e32 v9, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_or_b32_e32 v7, 0x400000, v2
-; VI-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; VI-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
-; VI-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
-; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; VI-NEXT: v_alignbit_b32 v2, v5, v2, 16
-; VI-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
-; VI-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; VI-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; VI-NEXT: s_cbranch_execnz .LBB16_1
-; VI-NEXT: ; %bb.2: ; %atomicrmw.end
-; VI-NEXT: s_or_b64 exec, exec, s[6:7]
-; VI-NEXT: v_mov_b32_e32 v0, v2
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: lds_atomic_fadd_ret_v2bf16:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: ds_read_b32 v2, v0
-; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v1
-; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB16_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v4, v2
-; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v4
-; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
-; GFX908-NEXT: v_add_f32_e32 v2, v2, v3
-; GFX908-NEXT: v_add_f32_e32 v5, v5, v1
-; GFX908-NEXT: v_bfe_u32 v6, v2, 16, 1
-; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v2
-; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
-; GFX908-NEXT: v_add3_u32 v6, v6, v2, s8
-; GFX908-NEXT: v_add3_u32 v8, v8, v5, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX908-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
-; GFX908-NEXT: v_perm_b32 v2, v5, v2, s9
-; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
-; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB16_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: v_mov_b32_e32 v0, v2
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: lds_atomic_fadd_ret_v2bf16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: ds_read_b32 v2, v0
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v1
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB16_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v4, v2
-; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v4
-; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
-; GFX90A-NEXT: v_add_f32_e32 v2, v2, v3
-; GFX90A-NEXT: v_add_f32_e32 v5, v5, v1
-; GFX90A-NEXT: v_bfe_u32 v6, v2, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v2
-; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
-; GFX90A-NEXT: v_add3_u32 v6, v6, v2, s8
-; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX90A-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
-; GFX90A-NEXT: v_perm_b32 v2, v5, v2, s9
-; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB16_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v2
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: lds_atomic_fadd_ret_v2bf16:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: ds_read_b32 v2, v0
-; GFX940-NEXT: s_mov_b64 s[2:3], 0
-; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v1
-; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB16_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v4, v2
-; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v4
-; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
-; GFX940-NEXT: v_add_f32_e32 v2, v2, v3
-; GFX940-NEXT: v_add_f32_e32 v5, v5, v1
-; GFX940-NEXT: v_bfe_u32 v6, v2, 16, 1
-; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v2
-; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
-; GFX940-NEXT: v_add3_u32 v6, v6, v2, s4
-; GFX940-NEXT: v_add3_u32 v8, v8, v5, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[0:1]
-; GFX940-NEXT: v_perm_b32 v2, v5, v2, s5
-; GFX940-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
-; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB16_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v0, v2
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: lds_atomic_fadd_ret_v2bf16:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: ds_load_b32 v2, v0
-; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v1
-; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB16_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v4, v2
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
-; GFX12-NEXT: v_add_f32_e32 v5, v5, v1
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT: v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
-; GFX12-NEXT: v_add_f32_e32 v2, v2, v3
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX12-NEXT: v_bfe_u32 v6, v2, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
-; GFX12-NEXT: v_add3_u32 v6, v6, v2, 0x7fff
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT: v_cndmask_b32_e64 v2, v6, v8, s0
-; GFX12-NEXT: v_perm_b32 v2, v5, v2, 0x7060302
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v4
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SE
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4
-; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB16_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: v_mov_b32_e32 v0, v2
-; GFX12-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX7-LABEL: lds_atomic_fadd_ret_v2bf16:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: s_mov_b32 m0, -1
-; GFX7-NEXT: ds_read_b32 v3, v0
-; GFX7-NEXT: v_mov_b32_e32 v4, v1
-; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
-; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
-; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX7-NEXT: .LBB16_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
-; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
-; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
-; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX7-NEXT: v_add_f32_e32 v5, v5, v4
-; GFX7-NEXT: v_alignbit_b32 v1, v1, v3, 16
-; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v6
-; GFX7-NEXT: v_alignbit_b32 v3, v3, v5, 16
-; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v0, v1, v3
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
-; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB16_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: v_mov_b32_e32 v0, v3
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: lds_atomic_fadd_ret_v2bf16:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: s_mov_b32 m0, -1
-; GFX8-NEXT: ds_read_b32 v3, v0
-; GFX8-NEXT: v_mov_b32_e32 v4, v1
-; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v4
-; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
-; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX8-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
-; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX8-NEXT: .LBB16_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GFX8-NEXT: v_mul_f32_e32 v3, 1.0, v3
-; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
-; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
-; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GFX8-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX8-NEXT: v_add_f32_e32 v5, v5, v4
-; GFX8-NEXT: v_alignbit_b32 v1, v1, v3, 16
-; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v6
-; GFX8-NEXT: v_alignbit_b32 v3, v3, v5, 16
-; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v0, v1, v3
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
-; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB16_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v0, v3
-; GFX8-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr addrspace(3) %ptr, <2 x bfloat> %val seq_cst
- ret <2 x bfloat> %result
-}
-
-define void @lds_atomic_fadd_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
-; VI-LABEL: lds_atomic_fadd_noret_v2bf16:
-; VI: ; %bb.0:
-; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NEXT: s_mov_b32 m0, -1
-; VI-NEXT: ds_read_b32 v3, v0
-; VI-NEXT: s_mov_b64 s[6:7], 0
-; VI-NEXT: v_lshlrev_b32_e32 v2, 16, v1
-; VI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; VI-NEXT: .LBB17_1: ; %atomicrmw.start
-; VI-NEXT: ; =>This Inner Loop Header: Depth=1
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v3
-; VI-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
-; VI-NEXT: v_add_f32_e32 v4, v4, v2
-; VI-NEXT: v_add_f32_e32 v5, v5, v1
-; VI-NEXT: v_bfe_u32 v6, v4, 16, 1
-; VI-NEXT: v_bfe_u32 v8, v5, 16, 1
-; VI-NEXT: v_add_u32_e32 v6, vcc, v6, v4
-; VI-NEXT: v_add_u32_e32 v8, vcc, v8, v5
-; VI-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
-; VI-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
-; VI-NEXT: v_or_b32_e32 v9, 0x400000, v5
-; VI-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; VI-NEXT: v_or_b32_e32 v7, 0x400000, v4
-; VI-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
-; VI-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
-; VI-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
-; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; VI-NEXT: v_alignbit_b32 v4, v5, v4, 16
-; VI-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
-; VI-NEXT: s_waitcnt lgkmcnt(0)
-; VI-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
-; VI-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; VI-NEXT: v_mov_b32_e32 v3, v4
-; VI-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; VI-NEXT: s_cbranch_execnz .LBB17_1
-; VI-NEXT: ; %bb.2: ; %atomicrmw.end
-; VI-NEXT: s_or_b64 exec, exec, s[6:7]
-; VI-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: lds_atomic_fadd_noret_v2bf16:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: ds_read_b32 v3, v0
-; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v1
-; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB17_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v3
-; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
-; GFX908-NEXT: v_add_f32_e32 v4, v4, v2
-; GFX908-NEXT: v_add_f32_e32 v5, v5, v1
-; GFX908-NEXT: v_bfe_u32 v6, v4, 16, 1
-; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v4
-; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
-; GFX908-NEXT: v_add3_u32 v6, v6, v4, s8
-; GFX908-NEXT: v_add3_u32 v8, v8, v5, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
-; GFX908-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
-; GFX908-NEXT: v_perm_b32 v4, v5, v4, s9
-; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
-; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX908-NEXT: v_mov_b32_e32 v3, v4
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB17_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: lds_atomic_fadd_noret_v2bf16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: ds_read_b32 v3, v0
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v1
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB17_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v3
-; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
-; GFX90A-NEXT: v_add_f32_e32 v4, v4, v2
-; GFX90A-NEXT: v_add_f32_e32 v5, v5, v1
-; GFX90A-NEXT: v_bfe_u32 v6, v4, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v4
-; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
-; GFX90A-NEXT: v_add3_u32 v6, v6, v4, s8
-; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4
-; GFX90A-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
-; GFX90A-NEXT: v_perm_b32 v4, v5, v4, s9
-; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB17_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: lds_atomic_fadd_noret_v2bf16:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: ds_read_b32 v3, v0
-; GFX940-NEXT: s_mov_b64 s[2:3], 0
-; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v1
-; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB17_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v3
-; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
-; GFX940-NEXT: v_add_f32_e32 v4, v4, v2
-; GFX940-NEXT: v_add_f32_e32 v5, v5, v1
-; GFX940-NEXT: v_bfe_u32 v6, v4, 16, 1
-; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v4
-; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
-; GFX940-NEXT: v_add3_u32 v6, v6, v4, s4
-; GFX940-NEXT: v_add3_u32 v8, v8, v5, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v4, v4
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[0:1]
-; GFX940-NEXT: v_perm_b32 v4, v5, v4, s5
-; GFX940-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
-; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB17_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: lds_atomic_fadd_noret_v2bf16:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: ds_load_b32 v3, v0
-; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v1
-; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB17_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT: v_dual_add_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
-; GFX12-NEXT: v_add_f32_e32 v4, v4, v2
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
-; GFX12-NEXT: v_bfe_u32 v6, v4, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v4
-; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
-; GFX12-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
-; GFX12-NEXT: v_cmp_u_f32_e64 s0, v4, v4
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
-; GFX12-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT: v_perm_b32 v4, v5, v4, 0x7060302
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
-; GFX12-NEXT: s_wait_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SE
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3
-; GFX12-NEXT: v_mov_b32_e32 v3, v4
-; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB17_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX7-LABEL: lds_atomic_fadd_noret_v2bf16:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: s_mov_b32 m0, -1
-; GFX7-NEXT: ds_read_b32 v4, v0
-; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
-; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX7-NEXT: .LBB17_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
-; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
-; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
-; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX7-NEXT: v_add_f32_e32 v5, v5, v1
-; GFX7-NEXT: v_alignbit_b32 v3, v3, v4, 16
-; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
-; GFX7-NEXT: v_alignbit_b32 v4, v4, v5, 16
-; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
-; GFX7-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
-; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB17_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: lds_atomic_fadd_noret_v2bf16:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: s_mov_b32 m0, -1
-; GFX8-NEXT: ds_read_b32 v4, v0
-; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
-; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX8-NEXT: .LBB17_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: v_mul_f32_e32 v3, 1.0, v3
-; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v4
-; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
-; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v3
-; GFX8-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX8-NEXT: v_add_f32_e32 v5, v5, v1
-; GFX8-NEXT: v_alignbit_b32 v3, v3, v4, 16
-; GFX8-NEXT: v_lshrrev_b32_e32 v4, 16, v6
-; GFX8-NEXT: v_alignbit_b32 v4, v4, v5, 16
-; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
-; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB17_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr addrspace(3) %ptr, <2 x bfloat> %val seq_cst
- ret void
-}
-
-!0 = !{}
>From f974c47983a7cb1f6ab798dfdbe0b99e1a7b3eff Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Tue, 11 Jun 2024 15:58:46 +0200
Subject: [PATCH 2/4] AMDGPU: Add codegen tests for global atomicrmw fmin/fmax
Continue working towards atomic test consistency.
---
.../global-atomicrmw-fadd-wrong-subtarget.ll | 75 +
.../CodeGen/AMDGPU/global-atomicrmw-fadd.ll | 16785 ++++++++++++++
.../CodeGen/AMDGPU/global-atomicrmw-fmax.ll | 18237 ++++++++++++++++
.../CodeGen/AMDGPU/global-atomicrmw-fmin.ll | 18237 ++++++++++++++++
.../CodeGen/AMDGPU/global-atomicrmw-fsub.ll | 17517 +++++++++++++++
llvm/test/CodeGen/AMDGPU/global-atomics-fp.ll | 8090 -------
6 files changed, 70851 insertions(+), 8090 deletions(-)
create mode 100644 llvm/test/CodeGen/AMDGPU/global-atomicrmw-fadd-wrong-subtarget.ll
create mode 100644 llvm/test/CodeGen/AMDGPU/global-atomicrmw-fadd.ll
create mode 100644 llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmax.ll
create mode 100644 llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmin.ll
create mode 100644 llvm/test/CodeGen/AMDGPU/global-atomicrmw-fsub.ll
delete mode 100644 llvm/test/CodeGen/AMDGPU/global-atomics-fp.ll
diff --git a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fadd-wrong-subtarget.ll b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fadd-wrong-subtarget.ll
new file mode 100644
index 0000000000000..9fc0b5c57cc3a
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fadd-wrong-subtarget.ll
@@ -0,0 +1,75 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=amdgcn-amd-amdpal < %s | FileCheck -check-prefix=GCN %s
+
+define amdgpu_kernel void @global_atomic_fadd_ret_f32_wrong_subtarget(ptr addrspace(1) %ptr) #1 {
+; GCN-LABEL: global_atomic_fadd_ret_f32_wrong_subtarget:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_mov_b64 s[4:5], exec
+; GCN-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
+; GCN-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GCN-NEXT: ; implicit-def: $vgpr1
+; GCN-NEXT: s_and_saveexec_b64 s[2:3], vcc
+; GCN-NEXT: s_cbranch_execz .LBB0_4
+; GCN-NEXT: ; %bb.1:
+; GCN-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GCN-NEXT: s_bcnt1_i32_b64 s7, s[4:5]
+; GCN-NEXT: v_cvt_f32_ubyte0_e32 v1, s7
+; GCN-NEXT: s_mov_b64 s[4:5], 0
+; GCN-NEXT: v_mul_f32_e32 v2, 4.0, v1
+; GCN-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-NEXT: s_load_dword s6, s[0:1], 0x0
+; GCN-NEXT: v_mov_b32_e32 v3, 0
+; GCN-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-NEXT: v_mov_b32_e32 v1, s6
+; GCN-NEXT: .LBB0_2: ; %atomicrmw.start
+; GCN-NEXT: ; =>This Inner Loop Header: Depth=1
+; GCN-NEXT: v_mov_b32_e32 v5, v1
+; GCN-NEXT: v_add_f32_e32 v4, v5, v2
+; GCN-NEXT: global_atomic_cmpswap v1, v3, v[4:5], s[0:1] glc
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_wbinvl1
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, v1, v5
+; GCN-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GCN-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GCN-NEXT: s_cbranch_execnz .LBB0_2
+; GCN-NEXT: ; %bb.3: ; %Flow
+; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
+; GCN-NEXT: .LBB0_4: ; %Flow2
+; GCN-NEXT: s_or_b64 exec, exec, s[2:3]
+; GCN-NEXT: v_readfirstlane_b32 s0, v1
+; GCN-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
+; GCN-NEXT: v_mad_f32 v0, v0, 4.0, s0
+; GCN-NEXT: global_store_dword v[0:1], v0, off
+; GCN-NEXT: s_endpgm
+ %result = atomicrmw fadd ptr addrspace(1) %ptr, float 4.0 syncscope("agent") seq_cst
+ store float %result, ptr addrspace(1) undef
+ ret void
+}
+
+define amdgpu_kernel void @global_atomic_fadd_noret_f32_wrong_subtarget(ptr addrspace(1) %ptr) #1 {
+; GCN-LABEL: global_atomic_fadd_noret_f32_wrong_subtarget:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_mov_b64 s[2:3], exec
+; GCN-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
+; GCN-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
+; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GCN-NEXT: s_cbranch_execz .LBB1_2
+; GCN-NEXT: ; %bb.1:
+; GCN-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GCN-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
+; GCN-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
+; GCN-NEXT: v_mov_b32_e32 v0, 0
+; GCN-NEXT: v_mul_f32_e32 v1, 4.0, v1
+; GCN-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: buffer_wbinvl1
+; GCN-NEXT: .LBB1_2:
+; GCN-NEXT: s_endpgm
+ %result = atomicrmw fadd ptr addrspace(1) %ptr, float 4.0 syncscope("agent") seq_cst
+ ret void
+}
+
+attributes #1 = { "denormal-fp-math-f32"="preserve-sign,preserve-sign" "target-cpu"="gfx803" "target-features"="+atomic-fadd-no-rtn-insts" "amdgpu-unsafe-fp-atomics"="true" }
diff --git a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fadd.ll b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fadd.ll
new file mode 100644
index 0000000000000..5f60c8ac2b3b7
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fadd.ll
@@ -0,0 +1,16785 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 < %s | FileCheck -check-prefix=GFX12 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx940 < %s | FileCheck -check-prefix=GFX940 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx90a < %s | FileCheck -check-prefix=GFX90A %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx908 < %s | FileCheck -check-prefix=GFX908 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tonga < %s | FileCheck -check-prefix=GFX8 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=hawaii < %s | FileCheck -check-prefix=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tahiti < %s | FileCheck -check-prefix=GFX6 %s
+
+; --------------------------------------------------------------------
+; float
+; --------------------------------------------------------------------
+
+define float @global_agent_atomic_fadd_ret_f32(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB0_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB0_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB0_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB0_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_f32:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB0_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @global_agent_atomic_fadd_ret_f32__offset12b_pos(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB1_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB1_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB1_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB1_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB1_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fadd ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @global_agent_atomic_fadd_ret_f32__offset12b_neg(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB2_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB2_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB2_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v0, v[3:4], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v3
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, -1, v4, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: v_add_f32_e32 v5, v6, v2
+; GFX7-NEXT: v_mov_b32_e32 v0, v5
+; GFX7-NEXT: v_mov_b32_e32 v1, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[0:1], v[3:4], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v0, v[3:4], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v3
+; GFX6-NEXT: v_addc_u32_e32 v4, vcc, -1, v4, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v0
+; GFX6-NEXT: v_add_f32_e32 v5, v6, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v0, v5
+; GFX6-NEXT: v_mov_b32_e32 v1, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], v[3:4], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB2_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 -512
+ %result = atomicrmw fadd ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define void @global_agent_atomic_fadd_noret_f32(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB3_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB3_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB3_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_f32:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB3_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fadd_noret_f32__offset12b_pos(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:2044
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:2044
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:2044
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB4_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:2044
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:2044
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB4_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB4_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB4_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fadd_noret_f32__offset12b_neg(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:-2048
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:-2048
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:-2048
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:-2048
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB5_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:-2048
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:-2048
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB5_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 -512
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define float @global_system_atomic_fadd_ret_f32__offset12b_pos(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_system_atomic_fadd_ret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB6_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fadd_ret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fadd_ret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB6_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fadd_ret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB6_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fadd_ret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB6_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fadd_ret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB6_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fadd_ret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB6_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fadd_ret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB6_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fadd_ret_f32__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB6_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fadd ptr addrspace(1) %gep, float %val seq_cst
+ ret float %result
+}
+
+define void @global_system_atomic_fadd_noret_f32__offset12b_pos(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_system_atomic_fadd_noret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB7_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fadd_noret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:2044 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fadd_noret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB7_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fadd_noret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB7_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fadd_noret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB7_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fadd_noret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB7_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fadd_noret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB7_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fadd_noret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB7_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fadd_noret_f32__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB7_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, float %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; float with ftz/daz
+; --------------------------------------------------------------------
+
+define float @global_agent_atomic_fadd_ret_f32__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_f32__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_f32__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_f32__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_f32__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB8_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_f32__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_f32__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_f32__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_f32__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB8_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_f32__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB8_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @global_agent_atomic_fadd_ret_f32__offset12b_pos__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB9_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB9_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB9_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB9_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB9_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fadd ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @global_agent_atomic_fadd_ret_f32__offset12b_neg__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_neg__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_neg__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_neg__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_neg__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB10_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_neg__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_neg__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB10_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_neg__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB10_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_neg__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v0, v[3:4], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v3
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, -1, v4, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: v_add_f32_e32 v5, v6, v2
+; GFX7-NEXT: v_mov_b32_e32 v0, v5
+; GFX7-NEXT: v_mov_b32_e32 v1, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[0:1], v[3:4], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB10_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_f32__offset12b_neg__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v0, v[3:4], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v3
+; GFX6-NEXT: v_addc_u32_e32 v4, vcc, -1, v4, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v0
+; GFX6-NEXT: v_add_f32_e32 v5, v6, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v0, v5
+; GFX6-NEXT: v_mov_b32_e32 v1, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], v[3:4], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB10_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 -512
+ %result = atomicrmw fadd ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define void @global_agent_atomic_fadd_noret_f32__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_f32__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_f32__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_f32__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_f32__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB11_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_f32__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_f32__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_f32__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_f32__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB11_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_f32__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB11_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fadd ptr addrspace(1) %ptr, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fadd_noret_f32__offset12b_pos__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:2044
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:2044
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:2044
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB12_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:2044
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:2044
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB12_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB12_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB12_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fadd_noret_f32__offset12b_neg__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_neg__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:-2048
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_neg__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:-2048
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_neg__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:-2048
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_neg__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:-2048
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB13_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_neg__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:-2048
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_neg__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:-2048
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_neg__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB13_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_neg__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB13_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_f32__offset12b_neg__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB13_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 -512
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define float @global_system_atomic_fadd_ret_f32__offset12b_pos__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_system_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB14_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_add_f32 v0, v[0:1], v2, off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB14_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB14_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB14_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fadd ptr addrspace(1) %gep, float %val seq_cst
+ ret float %result
+}
+
+define void @global_system_atomic_fadd_noret_f32__offset12b_pos__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_system_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB15_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_add_f32 v[0:1], v2, off offset:2044 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB15_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB15_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB15_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB15_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB15_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB15_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB15_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, float %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; double
+; --------------------------------------------------------------------
+
+define double @global_agent_atomic_fadd_ret_f64(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[4:5], v[0:1], off
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_add_f64_e32 v[4:5], v[6:7], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB16_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_add_f64 v[0:1], v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[4:5], v[0:1], off
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB16_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB16_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_atomic_add_f64 v[0:1], v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB16_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v7, v5
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB16_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: v_mov_b32_e32 v1, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_mov_b32_e32 v7, v1
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v[6:7], s[4:7], 0 addr64
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v11, v1
+; GFX7-NEXT: v_mov_b32_e32 v10, v0
+; GFX7-NEXT: v_add_f64 v[8:9], v[10:11], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v8
+; GFX7-NEXT: v_mov_b32_e32 v1, v9
+; GFX7-NEXT: v_mov_b32_e32 v2, v10
+; GFX7-NEXT: v_mov_b32_e32 v3, v11
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[6:7], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB16_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_f64:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_mov_b32_e32 v7, v1
+; GFX6-NEXT: v_mov_b32_e32 v6, v0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dwordx2 v[0:1], v[6:7], s[4:7], 0 addr64
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v1
+; GFX6-NEXT: v_mov_b32_e32 v10, v0
+; GFX6-NEXT: v_add_f64 v[8:9], v[10:11], v[4:5]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v0, v8
+; GFX6-NEXT: v_mov_b32_e32 v1, v9
+; GFX6-NEXT: v_mov_b32_e32 v2, v10
+; GFX6-NEXT: v_mov_b32_e32 v3, v11
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[6:7], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB16_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(1) %ptr, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define double @global_agent_atomic_fadd_ret_f64__offset12b_pos(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_f64__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[4:5], v[0:1], off offset:2040
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_add_f64_e32 v[4:5], v[6:7], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:2040 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB17_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_f64__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_add_f64 v[0:1], v[0:1], v[2:3], off offset:2040 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_f64__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[4:5], v[0:1], off offset:2040
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB17_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_f64__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB17_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_f64__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_atomic_add_f64 v[0:1], v[0:1], v[2:3], off offset:2040 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_f64__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_f64__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7f8, v0
+; GFX8-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dwordx2 v[0:1], v[4:5]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v9, v1
+; GFX8-NEXT: v_mov_b32_e32 v8, v0
+; GFX8-NEXT: v_add_f64 v[6:7], v[8:9], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_f64__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_mov_b32_e32 v7, v1
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v[6:7], s[4:7], 0 addr64 offset:2040
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v11, v1
+; GFX7-NEXT: v_mov_b32_e32 v10, v0
+; GFX7-NEXT: v_add_f64 v[8:9], v[10:11], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v8
+; GFX7-NEXT: v_mov_b32_e32 v1, v9
+; GFX7-NEXT: v_mov_b32_e32 v2, v10
+; GFX7-NEXT: v_mov_b32_e32 v3, v11
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[6:7], s[4:7], 0 addr64 offset:2040 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_f64__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_mov_b32_e32 v7, v1
+; GFX6-NEXT: v_mov_b32_e32 v6, v0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dwordx2 v[0:1], v[6:7], s[4:7], 0 addr64 offset:2040
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v1
+; GFX6-NEXT: v_mov_b32_e32 v10, v0
+; GFX6-NEXT: v_add_f64 v[8:9], v[10:11], v[4:5]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v0, v8
+; GFX6-NEXT: v_mov_b32_e32 v1, v9
+; GFX6-NEXT: v_mov_b32_e32 v2, v10
+; GFX6-NEXT: v_mov_b32_e32 v3, v11
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[6:7], s[4:7], 0 addr64 offset:2040 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB17_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(1) %ptr, i64 255
+ %result = atomicrmw fadd ptr addrspace(1) %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define double @global_agent_atomic_fadd_ret_f64__offset12b_neg(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_f64__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[4:5], v[0:1], off offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_add_f64_e32 v[4:5], v[6:7], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB18_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_f64__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_add_f64 v[0:1], v[0:1], v[2:3], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_f64__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[4:5], v[0:1], off offset:-2048
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB18_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_f64__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB18_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_f64__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_atomic_add_f64 v[0:1], v[0:1], v[2:3], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_f64__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB18_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_f64__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dwordx2 v[0:1], v[4:5]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v9, v1
+; GFX8-NEXT: v_mov_b32_e32 v8, v0
+; GFX8-NEXT: v_add_f64 v[6:7], v[8:9], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB18_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_f64__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: v_mov_b32_e32 v7, v1
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v[6:7], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v6, vcc, 0xfffff800, v6
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: v_addc_u32_e32 v7, vcc, -1, v7, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v11, v1
+; GFX7-NEXT: v_mov_b32_e32 v10, v0
+; GFX7-NEXT: v_add_f64 v[8:9], v[10:11], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v8
+; GFX7-NEXT: v_mov_b32_e32 v1, v9
+; GFX7-NEXT: v_mov_b32_e32 v2, v10
+; GFX7-NEXT: v_mov_b32_e32 v3, v11
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[6:7], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB18_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_f64__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: v_mov_b32_e32 v7, v1
+; GFX6-NEXT: v_mov_b32_e32 v6, v0
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dwordx2 v[0:1], v[6:7], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v6, vcc, 0xfffff800, v6
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: v_addc_u32_e32 v7, vcc, -1, v7, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v1
+; GFX6-NEXT: v_mov_b32_e32 v10, v0
+; GFX6-NEXT: v_add_f64 v[8:9], v[10:11], v[4:5]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v0, v8
+; GFX6-NEXT: v_mov_b32_e32 v1, v9
+; GFX6-NEXT: v_mov_b32_e32 v2, v10
+; GFX6-NEXT: v_mov_b32_e32 v3, v11
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[6:7], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB18_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(1) %ptr, i64 -256
+ %result = atomicrmw fadd ptr addrspace(1) %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define void @global_agent_atomic_fadd_noret_f64(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[6:7], v[0:1], off
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_add_f64_e32 v[4:5], v[6:7], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB19_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_add_f64 v[0:1], v[2:3], off
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[6:7], v[0:1], off
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB19_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[6:7], v[0:1], off
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB19_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_atomic_add_f64 v[0:1], v[2:3], off
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[6:7], v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB19_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dwordx2 v[6:7], v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v7, v5
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB19_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dwordx2 v[6:7], v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX7-NEXT: v_mov_b32_e32 v11, v7
+; GFX7-NEXT: v_mov_b32_e32 v10, v6
+; GFX7-NEXT: v_mov_b32_e32 v9, v5
+; GFX7-NEXT: v_mov_b32_e32 v8, v4
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v6, v8
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v7, v9
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB19_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_f64:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dwordx2 v[6:7], v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v7
+; GFX6-NEXT: v_mov_b32_e32 v10, v6
+; GFX6-NEXT: v_mov_b32_e32 v9, v5
+; GFX6-NEXT: v_mov_b32_e32 v8, v4
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v6, v8
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v7, v9
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB19_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fadd ptr addrspace(1) %ptr, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fadd_noret_f64__offset12b_pos(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_f64__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[6:7], v[0:1], off offset:2040
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_add_f64_e32 v[4:5], v[6:7], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:2040 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB20_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_f64__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_add_f64 v[0:1], v[2:3], off offset:2040
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_f64__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[6:7], v[0:1], off offset:2040
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB20_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_f64__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[6:7], v[0:1], off offset:2040
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB20_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_f64__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_atomic_add_f64 v[0:1], v[2:3], off offset:2040
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_f64__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[6:7], v[0:1], off offset:2040
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB20_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_f64__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7f8, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dwordx2 v[6:7], v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v7, v5
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB20_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_f64__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dwordx2 v[6:7], v[0:1], s[4:7], 0 addr64 offset:2040
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX7-NEXT: v_mov_b32_e32 v11, v7
+; GFX7-NEXT: v_mov_b32_e32 v10, v6
+; GFX7-NEXT: v_mov_b32_e32 v9, v5
+; GFX7-NEXT: v_mov_b32_e32 v8, v4
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 offset:2040 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v6, v8
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v7, v9
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB20_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_f64__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dwordx2 v[6:7], v[0:1], s[4:7], 0 addr64 offset:2040
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v7
+; GFX6-NEXT: v_mov_b32_e32 v10, v6
+; GFX6-NEXT: v_mov_b32_e32 v9, v5
+; GFX6-NEXT: v_mov_b32_e32 v8, v4
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 offset:2040 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v6, v8
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v7, v9
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB20_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(1) %ptr, i64 255
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fadd_noret_f64__offset12b_neg(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_f64__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[6:7], v[0:1], off offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_add_f64_e32 v[4:5], v[6:7], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB21_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_f64__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_add_f64 v[0:1], v[2:3], off offset:-2048
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_f64__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[6:7], v[0:1], off offset:-2048
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB21_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_f64__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[6:7], v[0:1], off offset:-2048
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB21_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_f64__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_atomic_add_f64 v[0:1], v[2:3], off offset:-2048
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_f64__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[6:7], v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB21_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_f64__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dwordx2 v[6:7], v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v7, v5
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB21_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_f64__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dwordx2 v[6:7], v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX7-NEXT: v_mov_b32_e32 v11, v7
+; GFX7-NEXT: v_mov_b32_e32 v10, v6
+; GFX7-NEXT: v_mov_b32_e32 v9, v5
+; GFX7-NEXT: v_mov_b32_e32 v8, v4
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v6, v8
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v7, v9
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB21_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_f64__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dwordx2 v[6:7], v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v7
+; GFX6-NEXT: v_mov_b32_e32 v10, v6
+; GFX6-NEXT: v_mov_b32_e32 v9, v5
+; GFX6-NEXT: v_mov_b32_e32 v8, v4
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v6, v8
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v7, v9
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB21_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(1) %ptr, i64 -256
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; half
+; --------------------------------------------------------------------
+
+define half @global_agent_atomic_fadd_ret_f16(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB22_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: global_load_dword v4, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX940-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB22_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB22_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB22_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v4, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX90A-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB22_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v4
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX908-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX908-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX908-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB22_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB22_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v5
+; GFX7-NEXT: v_not_b32_e32 v7, v2
+; GFX7-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v5, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v3, v4, v7
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v5, v2
+; GFX7-NEXT: v_or_b32_e32 v3, v3, v2
+; GFX7-NEXT: v_mov_b32_e32 v2, v3
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[2:3], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB22_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v5, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v5
+; GFX6-NEXT: v_not_b32_e32 v7, v2
+; GFX6-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v5, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v3, v4, v7
+; GFX6-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v5, v2
+; GFX6-NEXT: v_or_b32_e32 v3, v3, v2
+; GFX6-NEXT: v_mov_b32_e32 v2, v3
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[2:3], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB22_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v5, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(1) %ptr, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+define half @global_agent_atomic_fadd_ret_f16__offset12b_pos(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB23_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[6:7], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v6
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_load_dword v4, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX940-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB23_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB23_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB23_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v4, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX90A-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB23_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v4
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX908-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX908-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX908-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB23_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB23_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX7-NEXT: v_not_b32_e32 v8, v2
+; GFX7-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v7
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB23_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX6-NEXT: v_not_b32_e32 v8, v2
+; GFX6-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX6-NEXT: v_add_f32_e32 v2, v2, v7
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB23_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fadd ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+define half @global_agent_atomic_fadd_ret_f16__offset12b_neg(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB24_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[6:7], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v6
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_load_dword v4, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX940-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB24_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB24_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB24_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v4, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX90A-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB24_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v4
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX908-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX908-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX908-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB24_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB24_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX7-NEXT: v_not_b32_e32 v8, v2
+; GFX7-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v7
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB24_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX6-NEXT: v_not_b32_e32 v8, v2
+; GFX6-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX6-NEXT: v_add_f32_e32 v2, v2, v7
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB24_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 -1024
+ %result = atomicrmw fadd ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+ }
+
+define void @global_agent_atomic_fadd_noret_f16(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB25_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX940-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB25_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB25_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v4, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX10-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB25_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX90A-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB25_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX908-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB25_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX8-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB25_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: v_lshl_b32_e32 v3, 0xffff, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_not_b32_e32 v6, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v2, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX7-NEXT: v_add_f32_e32 v3, v3, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v2, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX7-NEXT: v_mov_b32_e32 v8, v4
+; GFX7-NEXT: v_mov_b32_e32 v7, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB25_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v2
+; GFX6-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX6-NEXT: v_lshl_b32_e32 v3, 0xffff, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_not_b32_e32 v6, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v2, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX6-NEXT: v_add_f32_e32 v3, v3, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v2, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX6-NEXT: v_mov_b32_e32 v8, v4
+; GFX6-NEXT: v_mov_b32_e32 v7, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB25_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fadd ptr addrspace(1) %ptr, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fadd_noret_f16__offset12b_pos(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB26_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX940-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB26_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB26_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v4, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX10-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB26_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX90A-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB26_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX908-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB26_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX8-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB26_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB26_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX6-NEXT: v_not_b32_e32 v6, v2
+; GFX6-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX6-NEXT: v_add_f32_e32 v2, v2, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB26_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fadd_noret_f16__offset12b_neg(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB27_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX940-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB27_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB27_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v4, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX10-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB27_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX90A-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB27_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX908-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB27_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX8-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB27_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB27_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_f16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX6-NEXT: v_not_b32_e32 v6, v2
+; GFX6-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX6-NEXT: v_add_f32_e32 v2, v2, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB27_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 -1024
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define half @global_agent_atomic_fadd_ret_f16__offset12b_pos__align4(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_pos__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f16_e32 v3, v4, v2
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB28_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_pos__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_add_f16_e32 v3, v5, v2
+; GFX940-NEXT: v_and_or_b32 v4, v5, s2, v3
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB28_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_pos__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f16_e32 v3, v4, v2
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB28_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_pos__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_add_f16_e32 v3, v4, v2
+; GFX10-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX10-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB28_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_pos__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_add_f16_e32 v3, v5, v2
+; GFX90A-NEXT: v_and_or_b32 v4, v5, s6, v3
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB28_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_pos__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_add_f16_e32 v3, v4, v2
+; GFX908-NEXT: v_and_or_b32 v3, v4, s6, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB28_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_pos__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_add_f16_e32 v0, v1, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX8-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB28_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_pos__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX7-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB28_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v4
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_f16__offset12b_pos__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX6-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX6-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB28_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fadd ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst, align 4
+ ret half %result
+}
+
+define void @global_agent_atomic_fadd_noret_f16__offset12b__align4_pos(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_f16__offset12b__align4_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off offset:2046
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_add_f16_e32 v3, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB29_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_f16__offset12b__align4_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v5, v[0:1], off offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_add_f16_e32 v3, v5, v2
+; GFX940-NEXT: v_and_or_b32 v4, v5, s2, v3
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB29_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_f16__offset12b__align4_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off offset:2046
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_add_f16_e32 v3, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB29_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_f16__offset12b__align4_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:2046
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f16_e32 v3, v4, v2
+; GFX10-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX10-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB29_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_f16__offset12b__align4_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f16_e32 v3, v5, v2
+; GFX90A-NEXT: v_and_or_b32 v4, v5, s6, v3
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB29_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_f16__offset12b__align4_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v4, v[0:1], off offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f16_e32 v3, v4, v2
+; GFX908-NEXT: v_and_or_b32 v3, v4, s6, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB29_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_f16__offset12b__align4_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f16_e32 v3, v4, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB29_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_f16__offset12b__align4_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v2
+; GFX7-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB29_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_f16__offset12b__align4_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v2
+; GFX6-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX6-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB29_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst, align 4
+ ret void
+}
+
+define half @global_system_atomic_fadd_ret_f16__offset12b_pos(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_system_atomic_fadd_ret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB30_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fadd_ret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[6:7], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v6
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_load_dword v4, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX940-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB30_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fadd_ret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB30_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fadd_ret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB30_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fadd_ret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v4, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX90A-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB30_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fadd_ret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v4
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX908-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX908-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX908-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB30_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fadd_ret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB30_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fadd_ret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX7-NEXT: v_not_b32_e32 v8, v2
+; GFX7-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v7
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB30_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fadd_ret_f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX6-NEXT: v_not_b32_e32 v8, v2
+; GFX6-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX6-NEXT: v_add_f32_e32 v2, v2, v7
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB30_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fadd ptr addrspace(1) %gep, half %val seq_cst
+ ret half %result
+}
+
+define void @global_system_atomic_fadd_noret_f16__offset12b_pos(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_system_atomic_fadd_noret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB31_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fadd_noret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX940-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB31_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fadd_noret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB31_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fadd_noret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v4, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX10-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB31_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fadd_noret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX90A-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB31_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fadd_noret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX908-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB31_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fadd_noret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX8-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB31_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fadd_noret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB31_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fadd_noret_f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX6-NEXT: v_not_b32_e32 v6, v2
+; GFX6-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX6-NEXT: v_add_f32_e32 v2, v2, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB31_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, half %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; bfloat
+; --------------------------------------------------------------------
+
+define bfloat @global_agent_atomic_fadd_ret_bf16(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB32_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB32_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB32_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB32_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB32_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB32_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB32_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v3, 0xffff, v5
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v6, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX7-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v5, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v7
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v3, v4, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v5, v2
+; GFX7-NEXT: v_or_b32_e32 v3, v3, v2
+; GFX7-NEXT: v_mov_b32_e32 v2, v3
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[2:3], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB32_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v5, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX6-NEXT: v_lshl_b32_e32 v3, 0xffff, v5
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v6, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX6-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v5, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_add_f32_e32 v2, v2, v7
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_and_b32_e32 v3, v4, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v5, v2
+; GFX6-NEXT: v_or_b32_e32 v3, v3, v2
+; GFX6-NEXT: v_mov_b32_e32 v2, v3
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[2:3], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB32_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v5, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(1) %ptr, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+define bfloat @global_agent_atomic_fadd_ret_bf16__offset12b_pos(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB33_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB33_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB33_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB33_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB33_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB33_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB33_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v7, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX7-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB33_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v7, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX6-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_add_f32_e32 v2, v2, v8
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB33_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fadd ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+define bfloat @global_agent_atomic_fadd_ret_bf16__offset12b_neg(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB34_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB34_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB34_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB34_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB34_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB34_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB34_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v7, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX7-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB34_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v7, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX6-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_add_f32_e32 v2, v2, v8
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB34_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 -1024
+ %result = atomicrmw fadd ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+ }
+
+define void @global_agent_atomic_fadd_noret_bf16(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB35_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v4, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v4, v4, v2
+; GFX940-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX940-NEXT: v_add3_u32 v7, v7, v4, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v4, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB35_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB35_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v4, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB35_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v4, v4, v2
+; GFX90A-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX90A-NEXT: v_add3_u32 v7, v7, v4, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB35_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB35_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX8-NEXT: v_bfe_u32 v8, v3, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v3
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB35_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v3, 0xffff, v5
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v6, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX7-NEXT: v_mov_b32_e32 v8, v4
+; GFX7-NEXT: v_mov_b32_e32 v7, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB35_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX6-NEXT: v_lshl_b32_e32 v3, 0xffff, v5
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v6, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX6-NEXT: v_mov_b32_e32 v8, v4
+; GFX6-NEXT: v_mov_b32_e32 v7, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB35_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fadd ptr addrspace(1) %ptr, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fadd_noret_bf16__offset12b_pos(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB36_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB36_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB36_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB36_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB36_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB36_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB36_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB36_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v5, v5
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB36_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fadd_noret_bf16__offset12b_neg(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB37_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB37_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB37_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB37_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB37_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB37_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB37_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB37_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v5, v5
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB37_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 -1024
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define bfloat @global_agent_atomic_fadd_ret_bf16__offset12b_pos__align4(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_pos__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX12-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB38_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_pos__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX940-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX940-NEXT: v_bfe_u32 v4, v3, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX940-NEXT: v_add3_u32 v4, v4, v3, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v3, v4, v6, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX940-NEXT: v_and_or_b32 v4, v5, s3, v3
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB38_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_pos__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX11-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB38_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_pos__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX10-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX10-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX10-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX10-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB38_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_pos__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX90A-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX90A-NEXT: v_bfe_u32 v4, v3, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX90A-NEXT: v_add3_u32 v4, v4, v3, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e32 v3, v4, v6, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX90A-NEXT: v_and_or_b32 v4, v5, s7, v3
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB38_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_pos__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX908-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX908-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, s7, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB38_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_pos__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v0
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX8-NEXT: v_cndmask_b32_e32 v0, v5, v7, vcc
+; GFX8-NEXT: v_or_b32_sdwa v5, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB38_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_pos__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB38_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v4
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_bf16__offset12b_pos__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX6-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB38_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fadd ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst, align 4
+ ret bfloat %result
+}
+
+define void @global_agent_atomic_fadd_noret_bf16__offset12b__align4_pos(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b__align4_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB39_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b__align4_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v5, v5, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: v_and_or_b32 v2, v3, s3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB39_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b__align4_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB39_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b__align4_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB39_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b__align4_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v5, v5, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v3, s7, v2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB39_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b__align4_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v5, v5, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: v_and_or_b32 v2, v3, s7, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB39_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b__align4_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v2
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v6, v7, vcc
+; GFX8-NEXT: v_or_b32_sdwa v2, v5, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB39_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b__align4_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX7-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB39_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b__align4_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX6-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX6-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB39_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst, align 4
+ ret void
+}
+
+define bfloat @global_system_atomic_fadd_ret_bf16__offset12b_pos(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_system_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB40_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB40_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB40_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB40_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB40_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB40_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB40_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v7, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX7-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB40_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v7, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX6-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_add_f32_e32 v2, v2, v8
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB40_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fadd ptr addrspace(1) %gep, bfloat %val seq_cst
+ ret bfloat %result
+}
+
+define void @global_system_atomic_fadd_noret_bf16__offset12b_pos(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_system_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB41_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB41_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB41_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB41_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB41_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB41_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB41_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB41_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v5, v5
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB41_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, bfloat %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x half>
+; --------------------------------------------------------------------
+
+define <2 x half> @global_agent_atomic_fadd_ret_v2f16(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB42_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB42_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB42_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB42_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB42_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB42_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_add_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v5, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB42_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX7-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX7-NEXT: v_mov_b32_e32 v9, v7
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB42_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_v2f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX6-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_add_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX6-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX6-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX6-NEXT: v_mov_b32_e32 v9, v7
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB42_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v2
+; GFX6-NEXT: v_mov_b32_e32 v1, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define <2 x half> @global_agent_atomic_fadd_ret_v2f16__offset12b_pos(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB43_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB43_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB43_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB43_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB43_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB43_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_add_f16_sdwa v0, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v5, v1, v2
+; GFX8-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB43_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX7-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX7-NEXT: v_mov_b32_e32 v9, v7
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB43_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX6-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_add_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX6-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX6-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX6-NEXT: v_mov_b32_e32 v9, v7
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB43_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v2
+; GFX6-NEXT: v_mov_b32_e32 v1, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define <2 x half> @global_agent_atomic_fadd_ret_v2f16__offset12b_neg(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_v2f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB44_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_v2f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB44_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_v2f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB44_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_v2f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB44_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_v2f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB44_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_v2f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB44_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_v2f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_add_f16_sdwa v0, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v5, v1, v2
+; GFX8-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB44_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_v2f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v6, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v2
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v0
+; GFX7-NEXT: v_mov_b32_e32 v9, v7
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[8:9], v[4:5], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB44_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_v2f16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v6, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v2
+; GFX6-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v6
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX6-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX6-NEXT: v_or_b32_e32 v6, v8, v0
+; GFX6-NEXT: v_mov_b32_e32 v9, v7
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[8:9], v[4:5], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB44_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 -512
+ %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define void @global_agent_atomic_fadd_noret_v2f16(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB45_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB45_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB45_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB45_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB45_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB45_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v5, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB45_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB45_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_v2f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX6-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX6-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB45_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fadd_noret_v2f16__offset12b_pos(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB46_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v5, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB46_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB46_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB46_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB46_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB46_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v5, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB46_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB46_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX6-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX6-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB46_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fadd_noret_v2f16__offset12b_neg(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_v2f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB47_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_v2f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v5, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB47_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_v2f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off offset:-2048
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB47_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_v2f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:-2048
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB47_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_v2f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB47_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_v2f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v4, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB47_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_v2f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v5, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB47_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_v2f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB47_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_v2f16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX6-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX6-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB47_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 -512
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x half> @global_system_atomic_fadd_ret_v2f16__offset12b_pos(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_system_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB48_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB48_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB48_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB48_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB48_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB48_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_add_f16_sdwa v0, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v5, v1, v2
+; GFX8-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB48_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX7-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX7-NEXT: v_mov_b32_e32 v9, v7
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB48_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX6-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_add_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX6-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX6-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX6-NEXT: v_mov_b32_e32 v9, v7
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB48_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v2
+; GFX6-NEXT: v_mov_b32_e32 v1, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x half> %val seq_cst
+ ret <2 x half> %result
+}
+
+define void @global_system_atomic_fadd_noret_v2f16__offset12b_pos(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_system_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB49_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v5, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB49_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB49_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB49_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB49_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB49_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v5, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB49_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB49_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX6-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX6-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB49_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, <2 x half> %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x bfloat>
+; --------------------------------------------------------------------
+
+define <2 x bfloat> @global_agent_atomic_fadd_ret_v2bf16(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB50_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB50_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB50_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB50_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB50_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB50_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v3
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v3, 16
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB50_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v2, v2, v6, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v3
+; GFX7-NEXT: v_mov_b32_e32 v6, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB50_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_v2bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX6-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_add_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v7
+; GFX6-NEXT: v_alignbit_b32 v2, v2, v6, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v3
+; GFX6-NEXT: v_mov_b32_e32 v6, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB50_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr addrspace(1) %ptr, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define <2 x bfloat> @global_agent_atomic_fadd_ret_v2bf16__offset12b_pos(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB51_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB51_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB51_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB51_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB51_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB51_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB51_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v2, v2, v6, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v3
+; GFX7-NEXT: v_mov_b32_e32 v6, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB51_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX6-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_add_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v7
+; GFX6-NEXT: v_alignbit_b32 v2, v2, v6, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v3
+; GFX6-NEXT: v_mov_b32_e32 v6, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB51_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define <2 x bfloat> @global_agent_atomic_fadd_ret_v2bf16__offset12b_neg(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_ret_v2bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB52_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_ret_v2bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB52_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_ret_v2bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB52_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_ret_v2bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB52_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_ret_v2bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB52_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_ret_v2bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB52_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_ret_v2bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB52_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_ret_v2bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v6, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v0
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v0, v0, v6, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v1
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[4:5], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB52_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_ret_v2bf16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v6, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v0
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX6-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v1
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v0
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 16, v7
+; GFX6-NEXT: v_alignbit_b32 v0, v0, v6, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v1
+; GFX6-NEXT: v_mov_b32_e32 v6, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[4:5], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v1
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB52_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 -512
+ %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @global_agent_atomic_fadd_noret_v2bf16(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB53_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB53_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB53_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB53_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB53_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB53_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB53_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB53_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_v2bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB53_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fadd ptr addrspace(1) %ptr, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fadd_noret_v2bf16__offset12b_pos(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB54_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB54_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB54_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB54_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB54_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB54_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB54_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB54_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB54_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fadd_noret_v2bf16__offset12b_neg(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fadd_noret_v2bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB55_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fadd_noret_v2bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB55_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fadd_noret_v2bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB55_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fadd_noret_v2bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB55_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fadd_noret_v2bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB55_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fadd_noret_v2bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB55_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fadd_noret_v2bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB55_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fadd_noret_v2bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB55_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fadd_noret_v2bf16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB55_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 -512
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x bfloat> @global_system_atomic_fadd_ret_v2bf16__offset12b_pos(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_system_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB56_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB56_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB56_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB56_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB56_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB56_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB56_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v2, v2, v6, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v3
+; GFX7-NEXT: v_mov_b32_e32 v6, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB56_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX6-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_add_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v7
+; GFX6-NEXT: v_alignbit_b32 v2, v2, v6, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v3
+; GFX6-NEXT: v_mov_b32_e32 v6, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB56_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x bfloat> %val seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @global_system_atomic_fadd_noret_v2bf16__offset12b_pos(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_system_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB57_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB57_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB57_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB57_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB57_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB57_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB57_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB57_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB57_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fadd ptr addrspace(1) %gep, <2 x bfloat> %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; misc
+; --------------------------------------------------------------------
+
+define amdgpu_kernel void @infer_as_before_atomic(ptr addrspace(4) %arg) #1 {
+; GFX12-LABEL: infer_as_before_atomic:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_mov_b32 s2, exec_lo
+; GFX12-NEXT: s_mov_b32 s3, exec_lo
+; GFX12-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_cmpx_eq_u32_e32 0, v0
+; GFX12-NEXT: s_cbranch_execz .LBB58_2
+; GFX12-NEXT: ; %bb.1:
+; GFX12-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
+; GFX12-NEXT: s_bcnt1_i32_b32 s2, s2
+; GFX12-NEXT: v_mov_b32_e32 v0, 0
+; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
+; GFX12-NEXT: .LBB58_2:
+; GFX12-NEXT: s_nop 0
+; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX12-NEXT: s_endpgm
+;
+; GFX940-LABEL: infer_as_before_atomic:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_mov_b64 s[2:3], exec
+; GFX940-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
+; GFX940-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX940-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX940-NEXT: s_cbranch_execz .LBB58_2
+; GFX940-NEXT: ; %bb.1:
+; GFX940-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX940-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, 0
+; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX940-NEXT: s_waitcnt lgkmcnt(0)
+; GFX940-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
+; GFX940-NEXT: .LBB58_2:
+; GFX940-NEXT: s_endpgm
+;
+; GFX11-LABEL: infer_as_before_atomic:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_mov_b32 s3, exec_lo
+; GFX11-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_cmpx_eq_u32_e32 0, v0
+; GFX11-NEXT: s_cbranch_execz .LBB58_2
+; GFX11-NEXT: ; %bb.1:
+; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
+; GFX11-NEXT: s_bcnt1_i32_b32 s2, s2
+; GFX11-NEXT: v_mov_b32_e32 v0, 0
+; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
+; GFX11-NEXT: .LBB58_2:
+; GFX11-NEXT: s_nop 0
+; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-NEXT: s_endpgm
+;
+; GFX10-LABEL: infer_as_before_atomic:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_mov_b32 s3, exec_lo
+; GFX10-NEXT: s_mov_b32 s2, 0
+; GFX10-NEXT: v_mbcnt_lo_u32_b32 v0, s3, 0
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
+; GFX10-NEXT: s_and_saveexec_b32 s4, vcc_lo
+; GFX10-NEXT: s_cbranch_execz .LBB58_3
+; GFX10-NEXT: ; %bb.1:
+; GFX10-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX10-NEXT: s_bcnt1_i32_b32 s3, s3
+; GFX10-NEXT: v_mov_b32_e32 v3, 0
+; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v2, s3
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: s_load_dword s4, s[0:1], 0x0
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, s4
+; GFX10-NEXT: .LBB58_2: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX10-NEXT: global_atomic_cmpswap v0, v3, v[0:1], s[0:1] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
+; GFX10-NEXT: s_cbranch_execnz .LBB58_2
+; GFX10-NEXT: .LBB58_3:
+; GFX10-NEXT: s_endpgm
+;
+; GFX90A-LABEL: infer_as_before_atomic:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_mov_b64 s[2:3], exec
+; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
+; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX90A-NEXT: s_cbranch_execz .LBB58_2
+; GFX90A-NEXT: ; %bb.1:
+; GFX90A-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX90A-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
+; GFX90A-NEXT: v_mov_b32_e32 v0, 0
+; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
+; GFX90A-NEXT: .LBB58_2:
+; GFX90A-NEXT: s_endpgm
+;
+; GFX908-LABEL: infer_as_before_atomic:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_mov_b64 s[2:3], exec
+; GFX908-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
+; GFX908-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX908-NEXT: s_cbranch_execz .LBB58_2
+; GFX908-NEXT: ; %bb.1:
+; GFX908-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX908-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
+; GFX908-NEXT: v_mov_b32_e32 v0, 0
+; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
+; GFX908-NEXT: .LBB58_2:
+; GFX908-NEXT: s_endpgm
+;
+; GFX8-LABEL: infer_as_before_atomic:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_mov_b64 s[2:3], exec
+; GFX8-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
+; GFX8-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX8-NEXT: s_cbranch_execz .LBB58_3
+; GFX8-NEXT: ; %bb.1:
+; GFX8-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX8-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
+; GFX8-NEXT: v_cvt_f32_ubyte0_e32 v4, s2
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX8-NEXT: s_mov_b64 s[0:1], 0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_load_dword s6, s[4:5], 0x0
+; GFX8-NEXT: v_mov_b32_e32 v0, s4
+; GFX8-NEXT: v_mov_b32_e32 v1, s5
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, s6
+; GFX8-NEXT: .LBB58_2: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: v_add_f32_e32 v2, v3, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX8-NEXT: s_cbranch_execnz .LBB58_2
+; GFX8-NEXT: .LBB58_3:
+; GFX8-NEXT: s_endpgm
+;
+; GFX7-LABEL: infer_as_before_atomic:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_mov_b64 s[2:3], exec
+; GFX7-NEXT: v_mbcnt_lo_u32_b32_e64 v0, s2, 0
+; GFX7-NEXT: v_mbcnt_hi_u32_b32_e32 v0, s3, v0
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX7-NEXT: s_cbranch_execz .LBB58_3
+; GFX7-NEXT: ; %bb.1:
+; GFX7-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX7-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: s_mov_b32 s3, 0xf000
+; GFX7-NEXT: v_cvt_f32_ubyte0_e32 v2, s2
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX7-NEXT: s_mov_b32 s2, -1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: s_load_dword s6, s[0:1], 0x0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v1, s6
+; GFX7-NEXT: .LBB58_2: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], off, s[0:3], 0 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB58_2
+; GFX7-NEXT: .LBB58_3:
+; GFX7-NEXT: s_endpgm
+;
+; GFX6-LABEL: infer_as_before_atomic:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_mov_b64 s[2:3], exec
+; GFX6-NEXT: v_mbcnt_lo_u32_b32_e64 v0, s2, 0
+; GFX6-NEXT: v_mbcnt_hi_u32_b32_e32 v0, s3, v0
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX6-NEXT: s_cbranch_execz .LBB58_3
+; GFX6-NEXT: ; %bb.1:
+; GFX6-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX6-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
+; GFX6-NEXT: s_mov_b64 s[4:5], 0
+; GFX6-NEXT: s_mov_b32 s3, 0xf000
+; GFX6-NEXT: v_cvt_f32_ubyte0_e32 v2, s2
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
+; GFX6-NEXT: s_mov_b32 s2, -1
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: s_load_dword s6, s[0:1], 0x0
+; GFX6-NEXT: s_waitcnt lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v1, s6
+; GFX6-NEXT: .LBB58_2: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], off, s[0:3], 0 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
+; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v1, v3
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB58_2
+; GFX6-NEXT: .LBB58_3:
+; GFX6-NEXT: s_endpgm
+ %load = load ptr, ptr addrspace(4) %arg
+ %v = atomicrmw fadd ptr %load, float 1.0 syncscope("agent-one-as") monotonic, align 4
+ ret void
+}
+
+attributes #0 = { nounwind "amdgpu-unsafe-fp-atomics"="true" }
+attributes #1 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" "amdgpu-unsafe-fp-atomics"="true" }
diff --git a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmax.ll b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmax.ll
new file mode 100644
index 0000000000000..492c74e56b367
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmax.ll
@@ -0,0 +1,18237 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 < %s | FileCheck -check-prefix=GFX12 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx940 < %s | FileCheck -check-prefix=GFX940 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx90a < %s | FileCheck -check-prefix=GFX90A %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx908 < %s | FileCheck -check-prefix=GFX908 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tonga < %s | FileCheck -check-prefix=GFX8 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=hawaii < %s | FileCheck -check-prefix=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tahiti < %s | FileCheck -check-prefix=GFX6 %s
+
+; --------------------------------------------------------------------
+; float
+; --------------------------------------------------------------------
+
+define float @global_agent_atomic_fmax_ret_f32(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB0_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB0_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB0_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX10-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB0_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB0_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB0_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX8-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_mul_f32_e32 v3, 1.0, v4
+; GFX8-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB0_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX7-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB0_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_f32:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX6-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB0_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(1) %ptr, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @global_agent_atomic_fmax_ret_f32__offset12b_pos(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB1_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB1_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB1_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX10-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB1_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB1_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB1_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB1_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX7-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB1_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX6-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB1_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fmax ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @global_agent_atomic_fmax_ret_f32__offset12b_neg(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB2_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB2_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB2_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX10-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB2_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB2_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB2_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB2_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v0, v[3:4], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v3
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, -1, v4, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX7-NEXT: v_max_f32_e32 v5, v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v0, v5
+; GFX7-NEXT: v_mov_b32_e32 v1, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[0:1], v[3:4], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v0, v[3:4], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v3
+; GFX6-NEXT: v_addc_u32_e32 v4, vcc, -1, v4, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX6-NEXT: v_max_f32_e32 v5, v0, v2
+; GFX6-NEXT: v_mov_b32_e32 v0, v5
+; GFX6-NEXT: v_mov_b32_e32 v1, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], v[3:4], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB2_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 -512
+ %result = atomicrmw fmax ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define void @global_agent_atomic_fmax_noret_f32(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB3_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB3_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB3_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB3_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB3_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB3_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB3_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB3_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_f32:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX6-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB3_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmax ptr addrspace(1) %ptr, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmax_noret_f32__offset12b_pos(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB4_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB4_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB4_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB4_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB4_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB4_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB4_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB4_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX6-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB4_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmax_noret_f32__offset12b_neg(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB5_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB5_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB5_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB5_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB5_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB5_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 -512
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define float @global_system_atomic_fmax_ret_f32__offset12b_pos(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_system_atomic_fmax_ret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB6_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmax_ret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB6_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmax_ret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB6_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmax_ret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX10-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB6_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmax_ret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB6_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmax_ret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB6_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmax_ret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB6_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmax_ret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX7-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB6_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmax_ret_f32__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX6-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB6_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fmax ptr addrspace(1) %gep, float %val seq_cst
+ ret float %result
+}
+
+define void @global_system_atomic_fmax_noret_f32__offset12b_pos(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_system_atomic_fmax_noret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB7_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmax_noret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB7_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmax_noret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB7_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmax_noret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB7_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmax_noret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB7_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmax_noret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB7_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmax_noret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB7_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmax_noret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB7_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmax_noret_f32__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX6-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB7_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, float %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; float with ftz/daz
+; --------------------------------------------------------------------
+
+define float @global_agent_atomic_fmax_ret_f32__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_f32__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB8_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_f32__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB8_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_f32__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB8_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_f32__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX10-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB8_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_f32__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB8_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_f32__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_f32__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX8-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_mul_f32_e32 v3, 1.0, v4
+; GFX8-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_f32__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX7-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB8_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_f32__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX6-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB8_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(1) %ptr, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @global_agent_atomic_fmax_ret_f32__offset12b_pos__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB9_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB9_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB9_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX10-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB9_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB9_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB9_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB9_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX7-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB9_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX6-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB9_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fmax ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @global_agent_atomic_fmax_ret_f32__offset12b_neg__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_neg__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB10_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_neg__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB10_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_neg__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB10_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_neg__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX10-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB10_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_neg__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB10_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_neg__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB10_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_neg__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB10_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_neg__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v0, v[3:4], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v3
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, -1, v4, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX7-NEXT: v_max_f32_e32 v5, v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v0, v5
+; GFX7-NEXT: v_mov_b32_e32 v1, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[0:1], v[3:4], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB10_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_f32__offset12b_neg__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v0, v[3:4], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v3
+; GFX6-NEXT: v_addc_u32_e32 v4, vcc, -1, v4, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX6-NEXT: v_max_f32_e32 v5, v0, v2
+; GFX6-NEXT: v_mov_b32_e32 v0, v5
+; GFX6-NEXT: v_mov_b32_e32 v1, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], v[3:4], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB10_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 -512
+ %result = atomicrmw fmax ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define void @global_agent_atomic_fmax_noret_f32__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_f32__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB11_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_f32__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB11_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_f32__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB11_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_f32__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB11_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_f32__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_f32__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_f32__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_f32__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB11_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_f32__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX6-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB11_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmax ptr addrspace(1) %ptr, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmax_noret_f32__offset12b_pos__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB12_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB12_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB12_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB12_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB12_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB12_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB12_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB12_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX6-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB12_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmax_noret_f32__offset12b_neg__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_neg__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB13_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_neg__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB13_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_neg__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB13_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_neg__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB13_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_neg__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB13_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_neg__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB13_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_neg__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB13_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_neg__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB13_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_f32__offset12b_neg__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB13_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 -512
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define float @global_system_atomic_fmax_ret_f32__offset12b_pos__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_system_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB14_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB14_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB14_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX10-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB14_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX7-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX6-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB14_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fmax ptr addrspace(1) %gep, float %val seq_cst
+ ret float %result
+}
+
+define void @global_system_atomic_fmax_noret_f32__offset12b_pos__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_system_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB15_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB15_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB15_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB15_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB15_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB15_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB15_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB15_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX6-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB15_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, float %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; double
+; --------------------------------------------------------------------
+
+define double @global_agent_atomic_fmax_ret_f64(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[4:5], v[0:1], off
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[6:7], v[6:7]
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[4:5], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB16_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX940-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB16_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[4:5], v[0:1], off
+; GFX11-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX11-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB16_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX10-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX10-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB16_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX90A-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX90A-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB16_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX908-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX908-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB16_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX8-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v7, v5
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX8-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB16_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: v_mov_b32_e32 v1, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_mov_b32_e32 v5, v1
+; GFX7-NEXT: v_mov_b32_e32 v4, v0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v[4:5], s[4:7], 0 addr64
+; GFX7-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v11, v1
+; GFX7-NEXT: v_mov_b32_e32 v10, v0
+; GFX7-NEXT: v_max_f64 v[0:1], v[10:11], v[10:11]
+; GFX7-NEXT: v_max_f64 v[8:9], v[0:1], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v0, v8
+; GFX7-NEXT: v_mov_b32_e32 v1, v9
+; GFX7-NEXT: v_mov_b32_e32 v2, v10
+; GFX7-NEXT: v_mov_b32_e32 v3, v11
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[4:5], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB16_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_f64:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_mov_b32_e32 v5, v1
+; GFX6-NEXT: v_mov_b32_e32 v4, v0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dwordx2 v[0:1], v[4:5], s[4:7], 0 addr64
+; GFX6-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v1
+; GFX6-NEXT: v_mov_b32_e32 v10, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_max_f64 v[0:1], v[10:11], v[10:11]
+; GFX6-NEXT: v_max_f64 v[8:9], v[0:1], v[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v0, v8
+; GFX6-NEXT: v_mov_b32_e32 v1, v9
+; GFX6-NEXT: v_mov_b32_e32 v2, v10
+; GFX6-NEXT: v_mov_b32_e32 v3, v11
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[4:5], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB16_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(1) %ptr, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define double @global_agent_atomic_fmax_ret_f64__offset12b_pos(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_f64__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[4:5], v[0:1], off offset:2040
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[6:7], v[6:7]
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[4:5], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:2040 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB17_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_f64__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX940-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:2040 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB17_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_f64__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[4:5], v[0:1], off offset:2040
+; GFX11-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX11-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB17_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_f64__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX10-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX10-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB17_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_f64__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX90A-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX90A-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB17_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_f64__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX908-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX908-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_f64__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7f8, v0
+; GFX8-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dwordx2 v[0:1], v[4:5]
+; GFX8-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v9, v1
+; GFX8-NEXT: v_mov_b32_e32 v8, v0
+; GFX8-NEXT: v_max_f64 v[0:1], v[8:9], v[8:9]
+; GFX8-NEXT: v_max_f64 v[6:7], v[0:1], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_f64__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_mov_b32_e32 v5, v1
+; GFX7-NEXT: v_mov_b32_e32 v4, v0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v[4:5], s[4:7], 0 addr64 offset:2040
+; GFX7-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v11, v1
+; GFX7-NEXT: v_mov_b32_e32 v10, v0
+; GFX7-NEXT: v_max_f64 v[0:1], v[10:11], v[10:11]
+; GFX7-NEXT: v_max_f64 v[8:9], v[0:1], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v0, v8
+; GFX7-NEXT: v_mov_b32_e32 v1, v9
+; GFX7-NEXT: v_mov_b32_e32 v2, v10
+; GFX7-NEXT: v_mov_b32_e32 v3, v11
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[4:5], s[4:7], 0 addr64 offset:2040 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_f64__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_mov_b32_e32 v5, v1
+; GFX6-NEXT: v_mov_b32_e32 v4, v0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dwordx2 v[0:1], v[4:5], s[4:7], 0 addr64 offset:2040
+; GFX6-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v1
+; GFX6-NEXT: v_mov_b32_e32 v10, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_max_f64 v[0:1], v[10:11], v[10:11]
+; GFX6-NEXT: v_max_f64 v[8:9], v[0:1], v[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v0, v8
+; GFX6-NEXT: v_mov_b32_e32 v1, v9
+; GFX6-NEXT: v_mov_b32_e32 v2, v10
+; GFX6-NEXT: v_mov_b32_e32 v3, v11
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[4:5], s[4:7], 0 addr64 offset:2040 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB17_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(1) %ptr, i64 255
+ %result = atomicrmw fmax ptr addrspace(1) %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define double @global_agent_atomic_fmax_ret_f64__offset12b_neg(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_f64__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[4:5], v[0:1], off offset:-2048
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[6:7], v[6:7]
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[4:5], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB18_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_f64__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX940-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB18_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_f64__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[4:5], v[0:1], off offset:-2048
+; GFX11-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX11-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB18_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_f64__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX10-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX10-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB18_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_f64__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX90A-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX90A-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB18_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_f64__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX908-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX908-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB18_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_f64__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dwordx2 v[0:1], v[4:5]
+; GFX8-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v9, v1
+; GFX8-NEXT: v_mov_b32_e32 v8, v0
+; GFX8-NEXT: v_max_f64 v[0:1], v[8:9], v[8:9]
+; GFX8-NEXT: v_max_f64 v[6:7], v[0:1], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB18_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_f64__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: v_mov_b32_e32 v5, v1
+; GFX7-NEXT: v_mov_b32_e32 v4, v0
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v[4:5], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v4
+; GFX7-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v5, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v11, v1
+; GFX7-NEXT: v_mov_b32_e32 v10, v0
+; GFX7-NEXT: v_max_f64 v[0:1], v[10:11], v[10:11]
+; GFX7-NEXT: v_max_f64 v[8:9], v[0:1], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v0, v8
+; GFX7-NEXT: v_mov_b32_e32 v1, v9
+; GFX7-NEXT: v_mov_b32_e32 v2, v10
+; GFX7-NEXT: v_mov_b32_e32 v3, v11
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[4:5], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB18_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_f64__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: v_mov_b32_e32 v5, v1
+; GFX6-NEXT: v_mov_b32_e32 v4, v0
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dwordx2 v[0:1], v[4:5], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v4
+; GFX6-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX6-NEXT: v_addc_u32_e32 v5, vcc, -1, v5, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v1
+; GFX6-NEXT: v_mov_b32_e32 v10, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_max_f64 v[0:1], v[10:11], v[10:11]
+; GFX6-NEXT: v_max_f64 v[8:9], v[0:1], v[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v0, v8
+; GFX6-NEXT: v_mov_b32_e32 v1, v9
+; GFX6-NEXT: v_mov_b32_e32 v2, v10
+; GFX6-NEXT: v_mov_b32_e32 v3, v11
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[4:5], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB18_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(1) %ptr, i64 -256
+ %result = atomicrmw fmax ptr addrspace(1) %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define void @global_agent_atomic_fmax_noret_f64(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[4:5], v[0:1], off
+; GFX12-NEXT: v_max_num_f64_e32 v[6:7], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[4:5], v[4:5]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[2:3], v[6:7]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX12-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB19_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX940-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB19_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[4:5], v[0:1], off
+; GFX11-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX11-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB19_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX10-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX10-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX10-NEXT: v_mov_b32_e32 v5, v3
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB19_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX90A-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX90A-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB19_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX908-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX908-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v5, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB19_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX8-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX8-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v5, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB19_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX7-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v11, v5
+; GFX7-NEXT: v_mov_b32_e32 v10, v4
+; GFX7-NEXT: v_mov_b32_e32 v9, v3
+; GFX7-NEXT: v_mov_b32_e32 v8, v2
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v8
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v5, v9
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB19_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_f64:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX6-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v5
+; GFX6-NEXT: v_mov_b32_e32 v10, v4
+; GFX6-NEXT: v_mov_b32_e32 v9, v3
+; GFX6-NEXT: v_mov_b32_e32 v8, v2
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v4, v8
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v5, v9
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB19_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmax ptr addrspace(1) %ptr, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmax_noret_f64__offset12b_pos(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_f64__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[4:5], v[0:1], off offset:2040
+; GFX12-NEXT: v_max_num_f64_e32 v[6:7], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[4:5], v[4:5]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[2:3], v[6:7]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5], off offset:2040 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX12-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB20_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_f64__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX940-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off offset:2040 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB20_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_f64__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[4:5], v[0:1], off offset:2040
+; GFX11-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5], off offset:2040 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX11-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB20_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_f64__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX10-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX10-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off offset:2040 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX10-NEXT: v_mov_b32_e32 v5, v3
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB20_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_f64__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX90A-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX90A-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off offset:2040 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB20_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_f64__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX908-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX908-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off offset:2040 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v5, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB20_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_f64__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7f8, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX8-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX8-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v5, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB20_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_f64__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64 offset:2040
+; GFX7-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX7-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v11, v5
+; GFX7-NEXT: v_mov_b32_e32 v10, v4
+; GFX7-NEXT: v_mov_b32_e32 v9, v3
+; GFX7-NEXT: v_mov_b32_e32 v8, v2
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 offset:2040 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v8
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v5, v9
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB20_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_f64__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64 offset:2040
+; GFX6-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX6-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v5
+; GFX6-NEXT: v_mov_b32_e32 v10, v4
+; GFX6-NEXT: v_mov_b32_e32 v9, v3
+; GFX6-NEXT: v_mov_b32_e32 v8, v2
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 offset:2040 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v4, v8
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v5, v9
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB20_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(1) %ptr, i64 255
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmax_noret_f64__offset12b_neg(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_f64__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[4:5], v[0:1], off offset:-2048
+; GFX12-NEXT: v_max_num_f64_e32 v[6:7], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[4:5], v[4:5]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[2:3], v[6:7]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX12-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB21_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_f64__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX940-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB21_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_f64__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[4:5], v[0:1], off offset:-2048
+; GFX11-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX11-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB21_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_f64__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX10-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX10-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX10-NEXT: v_mov_b32_e32 v5, v3
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB21_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_f64__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX90A-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX90A-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB21_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_f64__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX908-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX908-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v5, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB21_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_f64__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX8-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX8-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v5, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB21_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_f64__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX7-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v11, v5
+; GFX7-NEXT: v_mov_b32_e32 v10, v4
+; GFX7-NEXT: v_mov_b32_e32 v9, v3
+; GFX7-NEXT: v_mov_b32_e32 v8, v2
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v8
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v5, v9
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB21_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_f64__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX6-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v5
+; GFX6-NEXT: v_mov_b32_e32 v10, v4
+; GFX6-NEXT: v_mov_b32_e32 v9, v3
+; GFX6-NEXT: v_mov_b32_e32 v8, v2
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v4, v8
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v5, v9
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB21_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(1) %ptr, i64 -256
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; half
+; --------------------------------------------------------------------
+
+define half @global_agent_atomic_fmax_ret_f16(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v0
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB22_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB22_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB22_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB22_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB22_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB22_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB22_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v5
+; GFX7-NEXT: v_not_b32_e32 v7, v2
+; GFX7-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v5, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v3, v4, v7
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v5, v2
+; GFX7-NEXT: v_or_b32_e32 v3, v3, v2
+; GFX7-NEXT: v_mov_b32_e32 v2, v3
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[2:3], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB22_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v5, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v5
+; GFX6-NEXT: v_not_b32_e32 v7, v2
+; GFX6-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v5, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v3, v4, v7
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v5, v2
+; GFX6-NEXT: v_or_b32_e32 v3, v3, v2
+; GFX6-NEXT: v_mov_b32_e32 v2, v3
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[2:3], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB22_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v5, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(1) %ptr, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+define half @global_agent_atomic_fmax_ret_f16__offset12b_pos(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB23_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB23_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB23_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB23_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB23_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB23_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB23_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX7-NEXT: v_not_b32_e32 v8, v2
+; GFX7-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v7
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB23_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX6-NEXT: v_not_b32_e32 v8, v2
+; GFX6-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v7
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB23_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fmax ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+define half @global_agent_atomic_fmax_ret_f16__offset12b_neg(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB24_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB24_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB24_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB24_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB24_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB24_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB24_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX7-NEXT: v_not_b32_e32 v8, v2
+; GFX7-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v7
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB24_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX6-NEXT: v_not_b32_e32 v8, v2
+; GFX6-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v7
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB24_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 -1024
+ %result = atomicrmw fmax ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+ }
+
+define void @global_agent_atomic_fmax_noret_f16(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v0
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v3, v3, v3
+; GFX12-NEXT: v_max_num_f16_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB25_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX940-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX940-NEXT: v_max_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB25_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX11-NEXT: v_max_f16_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB25_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v4, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX10-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX10-NEXT: v_max_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB25_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX90A-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX90A-NEXT: v_max_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB25_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX908-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX908-NEXT: v_max_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB25_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX8-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX8-NEXT: v_max_f16_e32 v3, v3, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB25_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: v_lshl_b32_e32 v3, 0xffff, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_not_b32_e32 v6, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v2, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX7-NEXT: v_max_f32_e32 v3, v3, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v2, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX7-NEXT: v_mov_b32_e32 v8, v4
+; GFX7-NEXT: v_mov_b32_e32 v7, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB25_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v2
+; GFX6-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX6-NEXT: v_lshl_b32_e32 v3, 0xffff, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_not_b32_e32 v6, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v2, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX6-NEXT: v_max_f32_e32 v3, v3, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v2, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX6-NEXT: v_mov_b32_e32 v8, v4
+; GFX6-NEXT: v_mov_b32_e32 v7, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB25_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmax ptr addrspace(1) %ptr, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmax_noret_f16__offset12b_pos(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v6, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB26_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX940-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB26_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB26_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB26_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX90A-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB26_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX908-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB26_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX8-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB26_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB26_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX6-NEXT: v_not_b32_e32 v6, v2
+; GFX6-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB26_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmax_noret_f16__offset12b_neg(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v6, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB27_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX940-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB27_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB27_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB27_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX90A-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB27_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX908-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB27_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX8-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB27_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB27_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_f16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX6-NEXT: v_not_b32_e32 v6, v2
+; GFX6-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB27_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 -1024
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define half @global_agent_atomic_fmax_ret_f16__offset12b_pos__align4(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_pos__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v3, v4, v4
+; GFX12-NEXT: v_max_num_f16_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB28_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_pos__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f16_e32 v3, v5, v5
+; GFX940-NEXT: v_max_f16_e32 v3, v3, v2
+; GFX940-NEXT: v_and_or_b32 v4, v5, s2, v3
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB28_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_pos__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v3, v4, v4
+; GFX11-NEXT: v_max_f16_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB28_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_pos__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f16_e32 v3, v4, v4
+; GFX10-NEXT: v_max_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX10-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB28_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_pos__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f16_e32 v3, v5, v5
+; GFX90A-NEXT: v_max_f16_e32 v3, v3, v2
+; GFX90A-NEXT: v_and_or_b32 v4, v5, s6, v3
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB28_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_pos__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f16_e32 v3, v4, v4
+; GFX908-NEXT: v_max_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_and_or_b32 v3, v4, s6, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB28_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_pos__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v1, v2, v2
+; GFX8-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_max_f16_e32 v0, v6, v6
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX8-NEXT: v_max_f16_e32 v0, v0, v1
+; GFX8-NEXT: v_or_b32_e32 v5, v2, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB28_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_pos__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX7-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB28_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v4
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_f16__offset12b_pos__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX6-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB28_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fmax ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst, align 4
+ ret half %result
+}
+
+define void @global_agent_atomic_fmax_noret_f16__offset12b__align4_pos(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_f16__offset12b__align4_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX12-NEXT: v_max_num_f16_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f16_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v4
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB29_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_f16__offset12b__align4_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v4
+; GFX940-NEXT: v_and_or_b32 v2, v3, s2, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB29_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_f16__offset12b__align4_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX11-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v4
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB29_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_f16__offset12b__align4_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX10-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v4
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB29_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_f16__offset12b__align4_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v4
+; GFX90A-NEXT: v_and_or_b32 v2, v3, s6, v2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB29_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_f16__offset12b__align4_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v4
+; GFX908-NEXT: v_and_or_b32 v2, v3, s6, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB29_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_f16__offset12b__align4_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX8-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v4
+; GFX8-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB29_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_f16__offset12b__align4_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v2
+; GFX7-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB29_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_f16__offset12b__align4_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v2
+; GFX6-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB29_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst, align 4
+ ret void
+}
+
+define half @global_system_atomic_fmax_ret_f16__offset12b_pos(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_system_atomic_fmax_ret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB30_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmax_ret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB30_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmax_ret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB30_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmax_ret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB30_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmax_ret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB30_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmax_ret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB30_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmax_ret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB30_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmax_ret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX7-NEXT: v_not_b32_e32 v8, v2
+; GFX7-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v7
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB30_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmax_ret_f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX6-NEXT: v_not_b32_e32 v8, v2
+; GFX6-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v7
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB30_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fmax ptr addrspace(1) %gep, half %val seq_cst
+ ret half %result
+}
+
+define void @global_system_atomic_fmax_noret_f16__offset12b_pos(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_system_atomic_fmax_noret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v6, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB31_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmax_noret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX940-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB31_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmax_noret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB31_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmax_noret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB31_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmax_noret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX90A-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB31_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmax_noret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX908-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB31_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmax_noret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX8-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB31_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmax_noret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB31_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmax_noret_f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX6-NEXT: v_not_b32_e32 v6, v2
+; GFX6-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB31_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, half %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; bfloat
+; --------------------------------------------------------------------
+
+define bfloat @global_agent_atomic_fmax_ret_bf16(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB32_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB32_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB32_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB32_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB32_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB32_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB32_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v3, 0xffff, v5
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v6, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX7-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v5, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v7
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v3, v4, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v5, v2
+; GFX7-NEXT: v_or_b32_e32 v3, v3, v2
+; GFX7-NEXT: v_mov_b32_e32 v2, v3
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[2:3], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB32_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v5, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX6-NEXT: v_lshl_b32_e32 v3, 0xffff, v5
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v6, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX6-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v5, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v7
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_and_b32_e32 v3, v4, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v5, v2
+; GFX6-NEXT: v_or_b32_e32 v3, v3, v2
+; GFX6-NEXT: v_mov_b32_e32 v2, v3
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[2:3], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB32_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v5, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(1) %ptr, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+define bfloat @global_agent_atomic_fmax_ret_bf16__offset12b_pos(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB33_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB33_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB33_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB33_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB33_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB33_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB33_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v7, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX7-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB33_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v7, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX6-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v8
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB33_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fmax ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+define bfloat @global_agent_atomic_fmax_ret_bf16__offset12b_neg(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB34_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB34_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB34_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB34_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB34_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB34_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB34_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v7, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX7-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB34_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v7, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX6-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v8
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB34_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 -1024
+ %result = atomicrmw fmax ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+ }
+
+define void @global_agent_atomic_fmax_noret_bf16(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB35_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v4, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_max_f32_e32 v4, v4, v2
+; GFX940-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX940-NEXT: v_add3_u32 v7, v7, v4, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v4, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB35_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB35_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v4, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB35_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_max_f32_e32 v4, v4, v2
+; GFX90A-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX90A-NEXT: v_add3_u32 v7, v7, v4, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB35_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB35_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX8-NEXT: v_bfe_u32 v8, v3, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v3
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB35_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v3, 0xffff, v5
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v6, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX7-NEXT: v_mov_b32_e32 v8, v4
+; GFX7-NEXT: v_mov_b32_e32 v7, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB35_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX6-NEXT: v_lshl_b32_e32 v3, 0xffff, v5
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v6, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX6-NEXT: v_mov_b32_e32 v8, v4
+; GFX6-NEXT: v_mov_b32_e32 v7, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB35_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmax ptr addrspace(1) %ptr, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmax_noret_bf16__offset12b_pos(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB36_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB36_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB36_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB36_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB36_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB36_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB36_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB36_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v5, v5
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB36_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmax_noret_bf16__offset12b_neg(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB37_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB37_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB37_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB37_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB37_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB37_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB37_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB37_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v5, v5
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB37_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 -1024
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define bfloat @global_agent_atomic_fmax_ret_bf16__offset12b_pos__align4(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_pos__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB38_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_pos__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX940-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX940-NEXT: v_bfe_u32 v4, v3, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX940-NEXT: v_add3_u32 v4, v4, v3, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v3, v4, v6, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX940-NEXT: v_and_or_b32 v4, v5, s3, v3
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB38_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_pos__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB38_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_pos__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX10-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX10-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX10-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX10-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB38_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_pos__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX90A-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX90A-NEXT: v_bfe_u32 v4, v3, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX90A-NEXT: v_add3_u32 v4, v4, v3, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e32 v3, v4, v6, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX90A-NEXT: v_and_or_b32 v4, v5, s7, v3
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB38_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_pos__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX908-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, s7, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB38_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_pos__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_max_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v0
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX8-NEXT: v_cndmask_b32_e32 v0, v5, v7, vcc
+; GFX8-NEXT: v_or_b32_sdwa v5, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB38_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_pos__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB38_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v4
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_bf16__offset12b_pos__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB38_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fmax ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst, align 4
+ ret bfloat %result
+}
+
+define void @global_agent_atomic_fmax_noret_bf16__offset12b__align4_pos(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b__align4_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB39_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b__align4_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v5, v5, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: v_and_or_b32 v2, v3, s3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB39_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b__align4_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB39_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b__align4_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB39_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b__align4_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v5, v5, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v3, s7, v2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB39_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b__align4_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v5, v5, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: v_and_or_b32 v2, v3, s7, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB39_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b__align4_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v2
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v6, v7, vcc
+; GFX8-NEXT: v_or_b32_sdwa v2, v5, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB39_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b__align4_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX7-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB39_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b__align4_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX6-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB39_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst, align 4
+ ret void
+}
+
+define bfloat @global_system_atomic_fmax_ret_bf16__offset12b_pos(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_system_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB40_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB40_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB40_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB40_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB40_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB40_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB40_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v7, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX7-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB40_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v7, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX6-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v8
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB40_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fmax ptr addrspace(1) %gep, bfloat %val seq_cst
+ ret bfloat %result
+}
+
+define void @global_system_atomic_fmax_noret_bf16__offset12b_pos(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_system_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB41_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB41_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB41_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB41_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB41_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB41_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB41_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB41_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v5, v5
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB41_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, bfloat %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x half>
+; --------------------------------------------------------------------
+
+define <2 x half> @global_agent_atomic_fmax_ret_v2f16(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v3, v4, v4
+; GFX12-NEXT: v_pk_max_num_f16 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB42_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX940-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_max_f16 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB42_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX11-NEXT: v_pk_max_f16 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB42_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX10-NEXT: v_pk_max_f16 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB42_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX90A-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX90A-NEXT: v_pk_max_f16 v4, v3, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB42_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX908-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX908-NEXT: v_pk_max_f16 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB42_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v4, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v3
+; GFX8-NEXT: v_max_f16_sdwa v3, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v6, v6
+; GFX8-NEXT: v_max_f16_sdwa v3, v3, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB42_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX7-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX7-NEXT: v_mov_b32_e32 v9, v7
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB42_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_v2f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX6-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_max_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX6-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX6-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX6-NEXT: v_mov_b32_e32 v9, v7
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB42_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v2
+; GFX6-NEXT: v_mov_b32_e32 v1, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(1) %ptr, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define <2 x half> @global_agent_atomic_fmax_ret_v2f16__offset12b_pos(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v3, v4, v4
+; GFX12-NEXT: v_pk_max_num_f16 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB43_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX940-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_max_f16 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB43_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX11-NEXT: v_pk_max_f16 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB43_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX10-NEXT: v_pk_max_f16 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB43_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX90A-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX90A-NEXT: v_pk_max_f16 v4, v3, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB43_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX908-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX908-NEXT: v_pk_max_f16 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB43_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v1, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_max_f16_sdwa v0, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v6, v6
+; GFX8-NEXT: v_max_f16_sdwa v0, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB43_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX7-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX7-NEXT: v_mov_b32_e32 v9, v7
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB43_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX6-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_max_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX6-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX6-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX6-NEXT: v_mov_b32_e32 v9, v7
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB43_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v2
+; GFX6-NEXT: v_mov_b32_e32 v1, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fmax ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define <2 x half> @global_agent_atomic_fmax_ret_v2f16__offset12b_neg(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_v2f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v3, v4, v4
+; GFX12-NEXT: v_pk_max_num_f16 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB44_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_v2f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX940-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_max_f16 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB44_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_v2f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX11-NEXT: v_pk_max_f16 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB44_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_v2f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX10-NEXT: v_pk_max_f16 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB44_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_v2f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX90A-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX90A-NEXT: v_pk_max_f16 v4, v3, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB44_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_v2f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX908-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX908-NEXT: v_pk_max_f16 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB44_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_v2f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v1, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_max_f16_sdwa v0, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v6, v6
+; GFX8-NEXT: v_max_f16_sdwa v0, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB44_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_v2f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v6, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v2
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v0
+; GFX7-NEXT: v_mov_b32_e32 v9, v7
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[8:9], v[4:5], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB44_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_v2f16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v6, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v2
+; GFX6-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v6
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX6-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX6-NEXT: v_or_b32_e32 v6, v8, v0
+; GFX6-NEXT: v_mov_b32_e32 v9, v7
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[8:9], v[4:5], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB44_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 -512
+ %result = atomicrmw fmax ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define void @global_agent_atomic_fmax_noret_v2f16(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_pk_max_num_f16 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB45_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX940-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB45_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB45_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX10-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB45_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX90A-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB45_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX908-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB45_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v4, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v2, v2
+; GFX8-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v2, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v6, v3, v3
+; GFX8-NEXT: v_max_f16_sdwa v2, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v6, v6, v5
+; GFX8-NEXT: v_or_b32_e32 v2, v6, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB45_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB45_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_v2f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX6-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX6-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB45_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmax ptr addrspace(1) %ptr, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmax_noret_v2f16__offset12b_pos(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_pk_max_num_f16 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB46_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX940-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB46_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB46_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX10-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB46_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX90A-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB46_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX908-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB46_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v4, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v2, v2
+; GFX8-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v2, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v6, v3, v3
+; GFX8-NEXT: v_max_f16_sdwa v2, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v6, v6, v5
+; GFX8-NEXT: v_or_b32_e32 v2, v6, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB46_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB46_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX6-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX6-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB46_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmax_noret_v2f16__offset12b_neg(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_v2f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: v_pk_max_num_f16 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB47_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_v2f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX940-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB47_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_v2f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB47_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_v2f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX10-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB47_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_v2f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX90A-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB47_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_v2f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX908-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB47_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_v2f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v4, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v2, v2
+; GFX8-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v2, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v6, v3, v3
+; GFX8-NEXT: v_max_f16_sdwa v2, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v6, v6, v5
+; GFX8-NEXT: v_or_b32_e32 v2, v6, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB47_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_v2f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB47_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_v2f16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX6-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX6-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB47_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 -512
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x half> @global_system_atomic_fmax_ret_v2f16__offset12b_pos(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_system_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v3, v4, v4
+; GFX12-NEXT: v_pk_max_num_f16 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB48_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX940-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_max_f16 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB48_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX11-NEXT: v_pk_max_f16 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB48_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX10-NEXT: v_pk_max_f16 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB48_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX90A-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX90A-NEXT: v_pk_max_f16 v4, v3, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB48_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX908-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX908-NEXT: v_pk_max_f16 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB48_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v1, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_max_f16_sdwa v0, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v6, v6
+; GFX8-NEXT: v_max_f16_sdwa v0, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB48_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX7-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX7-NEXT: v_mov_b32_e32 v9, v7
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB48_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX6-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_max_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX6-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX6-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX6-NEXT: v_mov_b32_e32 v9, v7
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB48_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v2
+; GFX6-NEXT: v_mov_b32_e32 v1, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fmax ptr addrspace(1) %gep, <2 x half> %val seq_cst
+ ret <2 x half> %result
+}
+
+define void @global_system_atomic_fmax_noret_v2f16__offset12b_pos(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_system_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_pk_max_num_f16 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB49_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX940-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB49_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB49_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX10-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB49_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX90A-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB49_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX908-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB49_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v4, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v2, v2
+; GFX8-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v2, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v6, v3, v3
+; GFX8-NEXT: v_max_f16_sdwa v2, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v6, v6, v5
+; GFX8-NEXT: v_or_b32_e32 v2, v6, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB49_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB49_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX6-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX6-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB49_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, <2 x half> %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x bfloat>
+; --------------------------------------------------------------------
+
+define <2 x bfloat> @global_agent_atomic_fmax_ret_v2bf16(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_max_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB50_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB50_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB50_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB50_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB50_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB50_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v3
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v3, 16
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB50_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v2, v2, v7, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v3
+; GFX7-NEXT: v_mov_b32_e32 v6, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB50_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_v2bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX6-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_max_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v2, v2, v7, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v3
+; GFX6-NEXT: v_mov_b32_e32 v6, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB50_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr addrspace(1) %ptr, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define <2 x bfloat> @global_agent_atomic_fmax_ret_v2bf16__offset12b_pos(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_max_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB51_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB51_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB51_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB51_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB51_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB51_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_max_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB51_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v2, v2, v7, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v3
+; GFX7-NEXT: v_mov_b32_e32 v6, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB51_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX6-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_max_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v2, v2, v7, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v3
+; GFX6-NEXT: v_mov_b32_e32 v6, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB51_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fmax ptr addrspace(1) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define <2 x bfloat> @global_agent_atomic_fmax_ret_v2bf16__offset12b_neg(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_ret_v2bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_max_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB52_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_ret_v2bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB52_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_ret_v2bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB52_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_ret_v2bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB52_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_ret_v2bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB52_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_ret_v2bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB52_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_ret_v2bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_max_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB52_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_ret_v2bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v6, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v0
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v0, v0, v7, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v1
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[4:5], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB52_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_ret_v2bf16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v6, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v0
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX6-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v0
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v0, v0, v7, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v1
+; GFX6-NEXT: v_mov_b32_e32 v6, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[4:5], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v1
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB52_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 -512
+ %result = atomicrmw fmax ptr addrspace(1) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @global_agent_atomic_fmax_noret_v2bf16(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_max_num_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB53_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB53_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB53_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB53_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB53_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB53_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB53_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB53_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_v2bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB53_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmax ptr addrspace(1) %ptr, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmax_noret_v2bf16__offset12b_pos(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_max_num_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB54_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB54_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB54_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB54_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB54_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB54_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB54_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB54_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB54_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmax_noret_v2bf16__offset12b_neg(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmax_noret_v2bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_max_num_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB55_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmax_noret_v2bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB55_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmax_noret_v2bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB55_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmax_noret_v2bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB55_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmax_noret_v2bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB55_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmax_noret_v2bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB55_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmax_noret_v2bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB55_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmax_noret_v2bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB55_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmax_noret_v2bf16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB55_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 -512
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x bfloat> @global_system_atomic_fmax_ret_v2bf16__offset12b_pos(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_system_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_max_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB56_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB56_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB56_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB56_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB56_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB56_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_max_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB56_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v2, v2, v7, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v3
+; GFX7-NEXT: v_mov_b32_e32 v6, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB56_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX6-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_max_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v2, v2, v7, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v3
+; GFX6-NEXT: v_mov_b32_e32 v6, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB56_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fmax ptr addrspace(1) %gep, <2 x bfloat> %val seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @global_system_atomic_fmax_noret_v2bf16__offset12b_pos(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_system_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_max_num_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB57_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB57_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB57_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB57_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB57_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB57_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB57_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB57_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB57_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fmax ptr addrspace(1) %gep, <2 x bfloat> %val seq_cst
+ ret void
+}
+
+attributes #0 = { nounwind "amdgpu-unsafe-fp-atomics"="true" }
+attributes #1 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" "amdgpu-unsafe-fp-atomics"="true" }
diff --git a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmin.ll b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmin.ll
new file mode 100644
index 0000000000000..a4adb084da6bf
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmin.ll
@@ -0,0 +1,18237 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 < %s | FileCheck -check-prefix=GFX12 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx940 < %s | FileCheck -check-prefix=GFX940 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx90a < %s | FileCheck -check-prefix=GFX90A %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx908 < %s | FileCheck -check-prefix=GFX908 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tonga < %s | FileCheck -check-prefix=GFX8 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=hawaii < %s | FileCheck -check-prefix=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tahiti < %s | FileCheck -check-prefix=GFX6 %s
+
+; --------------------------------------------------------------------
+; float
+; --------------------------------------------------------------------
+
+define float @global_agent_atomic_fmin_ret_f32(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB0_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB0_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB0_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX10-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB0_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB0_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB0_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX8-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_mul_f32_e32 v3, 1.0, v4
+; GFX8-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB0_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX7-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB0_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_f32:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX6-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB0_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(1) %ptr, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @global_agent_atomic_fmin_ret_f32__offset12b_pos(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB1_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB1_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB1_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX10-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB1_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB1_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB1_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB1_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX7-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB1_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX6-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB1_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fmin ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @global_agent_atomic_fmin_ret_f32__offset12b_neg(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB2_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB2_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB2_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX10-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB2_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB2_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB2_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB2_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v0, v[3:4], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v3
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, -1, v4, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX7-NEXT: v_min_f32_e32 v5, v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v0, v5
+; GFX7-NEXT: v_mov_b32_e32 v1, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[0:1], v[3:4], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v0, v[3:4], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v3
+; GFX6-NEXT: v_addc_u32_e32 v4, vcc, -1, v4, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX6-NEXT: v_min_f32_e32 v5, v0, v2
+; GFX6-NEXT: v_mov_b32_e32 v0, v5
+; GFX6-NEXT: v_mov_b32_e32 v1, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], v[3:4], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB2_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 -512
+ %result = atomicrmw fmin ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define void @global_agent_atomic_fmin_noret_f32(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB3_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB3_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB3_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB3_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB3_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB3_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB3_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB3_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_f32:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX6-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB3_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmin ptr addrspace(1) %ptr, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmin_noret_f32__offset12b_pos(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB4_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB4_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB4_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB4_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB4_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB4_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB4_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB4_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX6-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB4_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmin_noret_f32__offset12b_neg(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB5_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB5_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB5_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB5_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB5_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB5_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 -512
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define float @global_system_atomic_fmin_ret_f32__offset12b_pos(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_system_atomic_fmin_ret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB6_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmin_ret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB6_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmin_ret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB6_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmin_ret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX10-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB6_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmin_ret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB6_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmin_ret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB6_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmin_ret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB6_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmin_ret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX7-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB6_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmin_ret_f32__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX6-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB6_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fmin ptr addrspace(1) %gep, float %val seq_cst
+ ret float %result
+}
+
+define void @global_system_atomic_fmin_noret_f32__offset12b_pos(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_system_atomic_fmin_noret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB7_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmin_noret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB7_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmin_noret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB7_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmin_noret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB7_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmin_noret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB7_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmin_noret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB7_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmin_noret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB7_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmin_noret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB7_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmin_noret_f32__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX6-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB7_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, float %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; float with ftz/daz
+; --------------------------------------------------------------------
+
+define float @global_agent_atomic_fmin_ret_f32__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_f32__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB8_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_f32__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB8_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_f32__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB8_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_f32__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX10-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB8_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_f32__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB8_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_f32__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_f32__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX8-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_mul_f32_e32 v3, 1.0, v4
+; GFX8-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_f32__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX7-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB8_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_f32__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX6-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB8_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(1) %ptr, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @global_agent_atomic_fmin_ret_f32__offset12b_pos__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB9_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB9_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB9_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX10-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB9_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB9_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB9_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB9_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX7-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB9_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX6-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB9_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fmin ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @global_agent_atomic_fmin_ret_f32__offset12b_neg__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_neg__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB10_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_neg__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB10_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_neg__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB10_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_neg__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX10-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB10_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_neg__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB10_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_neg__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB10_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_neg__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB10_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_neg__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v0, v[3:4], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v3
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, -1, v4, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX7-NEXT: v_min_f32_e32 v5, v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v0, v5
+; GFX7-NEXT: v_mov_b32_e32 v1, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[0:1], v[3:4], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB10_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_f32__offset12b_neg__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v0, v[3:4], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v3
+; GFX6-NEXT: v_addc_u32_e32 v4, vcc, -1, v4, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX6-NEXT: v_min_f32_e32 v5, v0, v2
+; GFX6-NEXT: v_mov_b32_e32 v0, v5
+; GFX6-NEXT: v_mov_b32_e32 v1, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], v[3:4], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB10_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 -512
+ %result = atomicrmw fmin ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define void @global_agent_atomic_fmin_noret_f32__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_f32__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB11_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_f32__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB11_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_f32__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB11_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_f32__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB11_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_f32__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_f32__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_f32__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_f32__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB11_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_f32__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX6-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB11_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmin ptr addrspace(1) %ptr, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmin_noret_f32__offset12b_pos__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB12_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB12_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB12_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB12_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB12_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB12_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB12_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB12_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX6-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB12_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmin_noret_f32__offset12b_neg__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_neg__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB13_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_neg__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB13_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_neg__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB13_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_neg__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB13_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_neg__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB13_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_neg__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB13_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_neg__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB13_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_neg__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB13_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_f32__offset12b_neg__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB13_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 -512
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define float @global_system_atomic_fmin_ret_f32__offset12b_pos__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_system_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB14_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB14_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB14_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX10-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB14_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX7-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v5
+; GFX6-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB14_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fmin ptr addrspace(1) %gep, float %val seq_cst
+ ret float %result
+}
+
+define void @global_system_atomic_fmin_noret_f32__offset12b_pos__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_system_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB15_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB15_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB15_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB15_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB15_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB15_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB15_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB15_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX6-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB15_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, float %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; double
+; --------------------------------------------------------------------
+
+define double @global_agent_atomic_fmin_ret_f64(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[4:5], v[0:1], off
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[6:7], v[6:7]
+; GFX12-NEXT: v_min_num_f64_e32 v[4:5], v[4:5], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB16_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX940-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB16_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[4:5], v[0:1], off
+; GFX11-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX11-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB16_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX10-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX10-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB16_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX90A-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX90A-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB16_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX908-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX908-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB16_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX8-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v7, v5
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX8-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB16_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: v_mov_b32_e32 v1, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_mov_b32_e32 v5, v1
+; GFX7-NEXT: v_mov_b32_e32 v4, v0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v[4:5], s[4:7], 0 addr64
+; GFX7-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v11, v1
+; GFX7-NEXT: v_mov_b32_e32 v10, v0
+; GFX7-NEXT: v_max_f64 v[0:1], v[10:11], v[10:11]
+; GFX7-NEXT: v_min_f64 v[8:9], v[0:1], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v0, v8
+; GFX7-NEXT: v_mov_b32_e32 v1, v9
+; GFX7-NEXT: v_mov_b32_e32 v2, v10
+; GFX7-NEXT: v_mov_b32_e32 v3, v11
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[4:5], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB16_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_f64:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_mov_b32_e32 v5, v1
+; GFX6-NEXT: v_mov_b32_e32 v4, v0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dwordx2 v[0:1], v[4:5], s[4:7], 0 addr64
+; GFX6-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v1
+; GFX6-NEXT: v_mov_b32_e32 v10, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_max_f64 v[0:1], v[10:11], v[10:11]
+; GFX6-NEXT: v_min_f64 v[8:9], v[0:1], v[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v0, v8
+; GFX6-NEXT: v_mov_b32_e32 v1, v9
+; GFX6-NEXT: v_mov_b32_e32 v2, v10
+; GFX6-NEXT: v_mov_b32_e32 v3, v11
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[4:5], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB16_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(1) %ptr, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define double @global_agent_atomic_fmin_ret_f64__offset12b_pos(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_f64__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[4:5], v[0:1], off offset:2040
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[6:7], v[6:7]
+; GFX12-NEXT: v_min_num_f64_e32 v[4:5], v[4:5], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:2040 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB17_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_f64__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX940-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:2040 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB17_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_f64__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[4:5], v[0:1], off offset:2040
+; GFX11-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX11-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB17_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_f64__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX10-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX10-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB17_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_f64__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX90A-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX90A-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB17_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_f64__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX908-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX908-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_f64__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7f8, v0
+; GFX8-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dwordx2 v[0:1], v[4:5]
+; GFX8-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v9, v1
+; GFX8-NEXT: v_mov_b32_e32 v8, v0
+; GFX8-NEXT: v_max_f64 v[0:1], v[8:9], v[8:9]
+; GFX8-NEXT: v_min_f64 v[6:7], v[0:1], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_f64__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_mov_b32_e32 v5, v1
+; GFX7-NEXT: v_mov_b32_e32 v4, v0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v[4:5], s[4:7], 0 addr64 offset:2040
+; GFX7-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v11, v1
+; GFX7-NEXT: v_mov_b32_e32 v10, v0
+; GFX7-NEXT: v_max_f64 v[0:1], v[10:11], v[10:11]
+; GFX7-NEXT: v_min_f64 v[8:9], v[0:1], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v0, v8
+; GFX7-NEXT: v_mov_b32_e32 v1, v9
+; GFX7-NEXT: v_mov_b32_e32 v2, v10
+; GFX7-NEXT: v_mov_b32_e32 v3, v11
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[4:5], s[4:7], 0 addr64 offset:2040 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_f64__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_mov_b32_e32 v5, v1
+; GFX6-NEXT: v_mov_b32_e32 v4, v0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dwordx2 v[0:1], v[4:5], s[4:7], 0 addr64 offset:2040
+; GFX6-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v1
+; GFX6-NEXT: v_mov_b32_e32 v10, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_max_f64 v[0:1], v[10:11], v[10:11]
+; GFX6-NEXT: v_min_f64 v[8:9], v[0:1], v[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v0, v8
+; GFX6-NEXT: v_mov_b32_e32 v1, v9
+; GFX6-NEXT: v_mov_b32_e32 v2, v10
+; GFX6-NEXT: v_mov_b32_e32 v3, v11
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[4:5], s[4:7], 0 addr64 offset:2040 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB17_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(1) %ptr, i64 255
+ %result = atomicrmw fmin ptr addrspace(1) %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define double @global_agent_atomic_fmin_ret_f64__offset12b_neg(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_f64__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[4:5], v[0:1], off offset:-2048
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[6:7], v[6:7]
+; GFX12-NEXT: v_min_num_f64_e32 v[4:5], v[4:5], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB18_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_f64__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX940-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB18_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_f64__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[4:5], v[0:1], off offset:-2048
+; GFX11-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX11-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB18_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_f64__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX10-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX10-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB18_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_f64__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX90A-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX90A-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB18_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_f64__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX908-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX908-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB18_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_f64__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dwordx2 v[0:1], v[4:5]
+; GFX8-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v9, v1
+; GFX8-NEXT: v_mov_b32_e32 v8, v0
+; GFX8-NEXT: v_max_f64 v[0:1], v[8:9], v[8:9]
+; GFX8-NEXT: v_min_f64 v[6:7], v[0:1], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB18_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_f64__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: v_mov_b32_e32 v5, v1
+; GFX7-NEXT: v_mov_b32_e32 v4, v0
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v[4:5], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v4
+; GFX7-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v5, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v11, v1
+; GFX7-NEXT: v_mov_b32_e32 v10, v0
+; GFX7-NEXT: v_max_f64 v[0:1], v[10:11], v[10:11]
+; GFX7-NEXT: v_min_f64 v[8:9], v[0:1], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v0, v8
+; GFX7-NEXT: v_mov_b32_e32 v1, v9
+; GFX7-NEXT: v_mov_b32_e32 v2, v10
+; GFX7-NEXT: v_mov_b32_e32 v3, v11
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[4:5], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB18_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_f64__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: v_mov_b32_e32 v5, v1
+; GFX6-NEXT: v_mov_b32_e32 v4, v0
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dwordx2 v[0:1], v[4:5], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v4
+; GFX6-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX6-NEXT: v_addc_u32_e32 v5, vcc, -1, v5, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v1
+; GFX6-NEXT: v_mov_b32_e32 v10, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_max_f64 v[0:1], v[10:11], v[10:11]
+; GFX6-NEXT: v_min_f64 v[8:9], v[0:1], v[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v0, v8
+; GFX6-NEXT: v_mov_b32_e32 v1, v9
+; GFX6-NEXT: v_mov_b32_e32 v2, v10
+; GFX6-NEXT: v_mov_b32_e32 v3, v11
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[4:5], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB18_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(1) %ptr, i64 -256
+ %result = atomicrmw fmin ptr addrspace(1) %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define void @global_agent_atomic_fmin_noret_f64(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[4:5], v[0:1], off
+; GFX12-NEXT: v_max_num_f64_e32 v[6:7], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[4:5], v[4:5]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f64_e32 v[2:3], v[2:3], v[6:7]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX12-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB19_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX940-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB19_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[4:5], v[0:1], off
+; GFX11-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX11-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB19_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX10-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX10-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX10-NEXT: v_mov_b32_e32 v5, v3
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB19_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX90A-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX90A-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB19_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX908-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX908-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v5, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB19_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX8-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX8-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v5, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB19_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX7-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v11, v5
+; GFX7-NEXT: v_mov_b32_e32 v10, v4
+; GFX7-NEXT: v_mov_b32_e32 v9, v3
+; GFX7-NEXT: v_mov_b32_e32 v8, v2
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v8
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v5, v9
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB19_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_f64:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX6-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v5
+; GFX6-NEXT: v_mov_b32_e32 v10, v4
+; GFX6-NEXT: v_mov_b32_e32 v9, v3
+; GFX6-NEXT: v_mov_b32_e32 v8, v2
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v4, v8
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v5, v9
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB19_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmin ptr addrspace(1) %ptr, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmin_noret_f64__offset12b_pos(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_f64__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[4:5], v[0:1], off offset:2040
+; GFX12-NEXT: v_max_num_f64_e32 v[6:7], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[4:5], v[4:5]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f64_e32 v[2:3], v[2:3], v[6:7]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5], off offset:2040 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX12-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB20_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_f64__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX940-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off offset:2040 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB20_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_f64__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[4:5], v[0:1], off offset:2040
+; GFX11-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5], off offset:2040 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX11-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB20_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_f64__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX10-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX10-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off offset:2040 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX10-NEXT: v_mov_b32_e32 v5, v3
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB20_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_f64__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX90A-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX90A-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off offset:2040 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB20_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_f64__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX908-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX908-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off offset:2040 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v5, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB20_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_f64__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7f8, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX8-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX8-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v5, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB20_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_f64__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64 offset:2040
+; GFX7-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX7-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v11, v5
+; GFX7-NEXT: v_mov_b32_e32 v10, v4
+; GFX7-NEXT: v_mov_b32_e32 v9, v3
+; GFX7-NEXT: v_mov_b32_e32 v8, v2
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 offset:2040 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v8
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v5, v9
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB20_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_f64__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64 offset:2040
+; GFX6-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX6-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v5
+; GFX6-NEXT: v_mov_b32_e32 v10, v4
+; GFX6-NEXT: v_mov_b32_e32 v9, v3
+; GFX6-NEXT: v_mov_b32_e32 v8, v2
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 offset:2040 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v4, v8
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v5, v9
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB20_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(1) %ptr, i64 255
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmin_noret_f64__offset12b_neg(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_f64__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[4:5], v[0:1], off offset:-2048
+; GFX12-NEXT: v_max_num_f64_e32 v[6:7], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[4:5], v[4:5]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f64_e32 v[2:3], v[2:3], v[6:7]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX12-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB21_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_f64__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX940-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB21_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_f64__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[4:5], v[0:1], off offset:-2048
+; GFX11-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX11-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB21_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_f64__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX10-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX10-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX10-NEXT: v_mov_b32_e32 v5, v3
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB21_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_f64__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX90A-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX90A-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB21_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_f64__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX908-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX908-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v5, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB21_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_f64__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX8-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX8-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v5, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB21_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_f64__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX7-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v11, v5
+; GFX7-NEXT: v_mov_b32_e32 v10, v4
+; GFX7-NEXT: v_mov_b32_e32 v9, v3
+; GFX7-NEXT: v_mov_b32_e32 v8, v2
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v8
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v5, v9
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB21_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_f64__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX6-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v5
+; GFX6-NEXT: v_mov_b32_e32 v10, v4
+; GFX6-NEXT: v_mov_b32_e32 v9, v3
+; GFX6-NEXT: v_mov_b32_e32 v8, v2
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v4, v8
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v5, v9
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB21_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(1) %ptr, i64 -256
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; half
+; --------------------------------------------------------------------
+
+define half @global_agent_atomic_fmin_ret_f16(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v0
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f16_e32 v5, v5, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB22_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX940-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB22_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB22_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX10-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB22_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX90A-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB22_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX908-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB22_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB22_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v5
+; GFX7-NEXT: v_not_b32_e32 v7, v2
+; GFX7-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v5, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v3, v4, v7
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v5, v2
+; GFX7-NEXT: v_or_b32_e32 v3, v3, v2
+; GFX7-NEXT: v_mov_b32_e32 v2, v3
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[2:3], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB22_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v5, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v5
+; GFX6-NEXT: v_not_b32_e32 v7, v2
+; GFX6-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v5, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v3, v4, v7
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v5, v2
+; GFX6-NEXT: v_or_b32_e32 v3, v3, v2
+; GFX6-NEXT: v_mov_b32_e32 v2, v3
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[2:3], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB22_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v5, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(1) %ptr, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+define half @global_agent_atomic_fmin_ret_f16__offset12b_pos(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f16_e32 v5, v5, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB23_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX940-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB23_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB23_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX10-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB23_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX90A-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB23_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX908-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB23_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB23_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX7-NEXT: v_not_b32_e32 v8, v2
+; GFX7-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v7
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB23_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX6-NEXT: v_not_b32_e32 v8, v2
+; GFX6-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v7
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB23_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fmin ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+define half @global_agent_atomic_fmin_ret_f16__offset12b_neg(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f16_e32 v5, v5, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB24_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX940-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB24_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB24_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX10-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB24_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX90A-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB24_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX908-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB24_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB24_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX7-NEXT: v_not_b32_e32 v8, v2
+; GFX7-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v7
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB24_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX6-NEXT: v_not_b32_e32 v8, v2
+; GFX6-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v7
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB24_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 -1024
+ %result = atomicrmw fmin ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+ }
+
+define void @global_agent_atomic_fmin_noret_f16(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v0
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v3, v3, v3
+; GFX12-NEXT: v_min_num_f16_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB25_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX940-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX940-NEXT: v_min_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB25_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX11-NEXT: v_min_f16_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB25_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v4, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX10-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX10-NEXT: v_min_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB25_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX90A-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX90A-NEXT: v_min_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB25_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX908-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX908-NEXT: v_min_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB25_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX8-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX8-NEXT: v_min_f16_e32 v3, v3, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB25_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: v_lshl_b32_e32 v3, 0xffff, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_not_b32_e32 v6, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v2, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX7-NEXT: v_min_f32_e32 v3, v3, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v2, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX7-NEXT: v_mov_b32_e32 v8, v4
+; GFX7-NEXT: v_mov_b32_e32 v7, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB25_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v2
+; GFX6-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX6-NEXT: v_lshl_b32_e32 v3, 0xffff, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_not_b32_e32 v6, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v2, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX6-NEXT: v_min_f32_e32 v3, v3, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v2, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX6-NEXT: v_mov_b32_e32 v8, v4
+; GFX6-NEXT: v_mov_b32_e32 v7, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB25_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmin ptr addrspace(1) %ptr, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmin_noret_f16__offset12b_pos(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v6, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: v_min_num_f16_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB26_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX940-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB26_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB26_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB26_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX90A-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB26_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX908-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB26_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX8-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB26_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB26_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX6-NEXT: v_not_b32_e32 v6, v2
+; GFX6-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB26_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmin_noret_f16__offset12b_neg(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v6, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: v_min_num_f16_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB27_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX940-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB27_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB27_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB27_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX90A-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB27_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX908-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB27_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX8-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB27_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB27_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_f16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX6-NEXT: v_not_b32_e32 v6, v2
+; GFX6-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB27_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 -1024
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define half @global_agent_atomic_fmin_ret_f16__offset12b_pos__align4(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_pos__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v3, v4, v4
+; GFX12-NEXT: v_min_num_f16_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB28_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_pos__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f16_e32 v3, v5, v5
+; GFX940-NEXT: v_min_f16_e32 v3, v3, v2
+; GFX940-NEXT: v_and_or_b32 v4, v5, s2, v3
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB28_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_pos__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v3, v4, v4
+; GFX11-NEXT: v_min_f16_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB28_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_pos__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f16_e32 v3, v4, v4
+; GFX10-NEXT: v_min_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX10-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB28_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_pos__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f16_e32 v3, v5, v5
+; GFX90A-NEXT: v_min_f16_e32 v3, v3, v2
+; GFX90A-NEXT: v_and_or_b32 v4, v5, s6, v3
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB28_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_pos__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f16_e32 v3, v4, v4
+; GFX908-NEXT: v_min_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_and_or_b32 v3, v4, s6, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB28_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_pos__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v1, v2, v2
+; GFX8-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_max_f16_e32 v0, v6, v6
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX8-NEXT: v_min_f16_e32 v0, v0, v1
+; GFX8-NEXT: v_or_b32_e32 v5, v2, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB28_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_pos__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX7-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB28_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v4
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_f16__offset12b_pos__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX6-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB28_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fmin ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst, align 4
+ ret half %result
+}
+
+define void @global_agent_atomic_fmin_noret_f16__offset12b__align4_pos(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_f16__offset12b__align4_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX12-NEXT: v_max_num_f16_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f16_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f16_e32 v2, v2, v4
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB29_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_f16__offset12b__align4_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX940-NEXT: v_min_f16_e32 v2, v2, v4
+; GFX940-NEXT: v_and_or_b32 v2, v3, s2, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB29_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_f16__offset12b__align4_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX11-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f16_e32 v2, v2, v4
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB29_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_f16__offset12b__align4_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX10-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX10-NEXT: v_min_f16_e32 v2, v2, v4
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB29_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_f16__offset12b__align4_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX90A-NEXT: v_min_f16_e32 v2, v2, v4
+; GFX90A-NEXT: v_and_or_b32 v2, v3, s6, v2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB29_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_f16__offset12b__align4_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX908-NEXT: v_min_f16_e32 v2, v2, v4
+; GFX908-NEXT: v_and_or_b32 v2, v3, s6, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB29_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_f16__offset12b__align4_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX8-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX8-NEXT: v_min_f16_e32 v2, v2, v4
+; GFX8-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB29_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_f16__offset12b__align4_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v2
+; GFX7-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB29_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_f16__offset12b__align4_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v2
+; GFX6-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB29_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst, align 4
+ ret void
+}
+
+define half @global_system_atomic_fmin_ret_f16__offset12b_pos(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_system_atomic_fmin_ret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f16_e32 v5, v5, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB30_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmin_ret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX940-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB30_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmin_ret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB30_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmin_ret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX10-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB30_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmin_ret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX90A-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB30_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmin_ret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX908-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB30_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmin_ret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB30_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmin_ret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX7-NEXT: v_not_b32_e32 v8, v2
+; GFX7-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v7
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB30_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmin_ret_f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX6-NEXT: v_not_b32_e32 v8, v2
+; GFX6-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v7
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB30_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fmin ptr addrspace(1) %gep, half %val seq_cst
+ ret half %result
+}
+
+define void @global_system_atomic_fmin_noret_f16__offset12b_pos(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_system_atomic_fmin_noret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v6, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: v_min_num_f16_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB31_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmin_noret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX940-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB31_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmin_noret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB31_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmin_noret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB31_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmin_noret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX90A-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB31_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmin_noret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX908-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB31_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmin_noret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX8-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB31_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmin_noret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB31_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmin_noret_f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX6-NEXT: v_not_b32_e32 v6, v2
+; GFX6-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB31_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, half %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; bfloat
+; --------------------------------------------------------------------
+
+define bfloat @global_agent_atomic_fmin_ret_bf16(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB32_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB32_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB32_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB32_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB32_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB32_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB32_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v3, 0xffff, v5
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v6, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX7-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v5, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v7
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v3, v4, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v5, v2
+; GFX7-NEXT: v_or_b32_e32 v3, v3, v2
+; GFX7-NEXT: v_mov_b32_e32 v2, v3
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[2:3], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB32_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v5, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX6-NEXT: v_lshl_b32_e32 v3, 0xffff, v5
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v6, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX6-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v5, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v7
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_and_b32_e32 v3, v4, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v5, v2
+; GFX6-NEXT: v_or_b32_e32 v3, v3, v2
+; GFX6-NEXT: v_mov_b32_e32 v2, v3
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[2:3], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB32_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v5, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(1) %ptr, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+define bfloat @global_agent_atomic_fmin_ret_bf16__offset12b_pos(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB33_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB33_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB33_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB33_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB33_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB33_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB33_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v7, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX7-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB33_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v7, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX6-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v8
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB33_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fmin ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+define bfloat @global_agent_atomic_fmin_ret_bf16__offset12b_neg(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB34_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB34_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB34_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB34_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB34_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB34_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB34_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v7, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX7-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB34_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v7, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX6-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v8
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB34_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 -1024
+ %result = atomicrmw fmin ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+ }
+
+define void @global_agent_atomic_fmin_noret_bf16(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB35_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v4, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_min_f32_e32 v4, v4, v2
+; GFX940-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX940-NEXT: v_add3_u32 v7, v7, v4, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v4, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB35_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB35_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v4, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB35_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_min_f32_e32 v4, v4, v2
+; GFX90A-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX90A-NEXT: v_add3_u32 v7, v7, v4, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB35_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB35_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX8-NEXT: v_bfe_u32 v8, v3, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v3
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB35_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v3, 0xffff, v5
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v6, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX7-NEXT: v_mov_b32_e32 v8, v4
+; GFX7-NEXT: v_mov_b32_e32 v7, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB35_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX6-NEXT: v_lshl_b32_e32 v3, 0xffff, v5
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v6, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX6-NEXT: v_mov_b32_e32 v8, v4
+; GFX6-NEXT: v_mov_b32_e32 v7, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB35_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmin ptr addrspace(1) %ptr, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmin_noret_bf16__offset12b_pos(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB36_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB36_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB36_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB36_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB36_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB36_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB36_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB36_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v5, v5
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB36_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmin_noret_bf16__offset12b_neg(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB37_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB37_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB37_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB37_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB37_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB37_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB37_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB37_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v5, v5
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB37_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 -1024
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define bfloat @global_agent_atomic_fmin_ret_bf16__offset12b_pos__align4(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_pos__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB38_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_pos__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX940-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX940-NEXT: v_bfe_u32 v4, v3, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX940-NEXT: v_add3_u32 v4, v4, v3, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v3, v4, v6, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX940-NEXT: v_and_or_b32 v4, v5, s3, v3
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB38_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_pos__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB38_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_pos__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX10-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX10-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX10-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX10-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB38_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_pos__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX90A-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX90A-NEXT: v_bfe_u32 v4, v3, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX90A-NEXT: v_add3_u32 v4, v4, v3, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e32 v3, v4, v6, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX90A-NEXT: v_and_or_b32 v4, v5, s7, v3
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB38_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_pos__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX908-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, s7, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB38_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_pos__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_min_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v0
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX8-NEXT: v_cndmask_b32_e32 v0, v5, v7, vcc
+; GFX8-NEXT: v_or_b32_sdwa v5, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB38_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_pos__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB38_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v4
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_bf16__offset12b_pos__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB38_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fmin ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst, align 4
+ ret bfloat %result
+}
+
+define void @global_agent_atomic_fmin_noret_bf16__offset12b__align4_pos(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b__align4_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB39_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b__align4_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v5, v5, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: v_and_or_b32 v2, v3, s3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB39_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b__align4_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB39_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b__align4_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB39_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b__align4_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v5, v5, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v3, s7, v2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB39_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b__align4_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v5, v5, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: v_and_or_b32 v2, v3, s7, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB39_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b__align4_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v2
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v6, v7, vcc
+; GFX8-NEXT: v_or_b32_sdwa v2, v5, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB39_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b__align4_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX7-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB39_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b__align4_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX6-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB39_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst, align 4
+ ret void
+}
+
+define bfloat @global_system_atomic_fmin_ret_bf16__offset12b_pos(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_system_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB40_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB40_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB40_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB40_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB40_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB40_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB40_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v7, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX7-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB40_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v7, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX6-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v8
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB40_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fmin ptr addrspace(1) %gep, bfloat %val seq_cst
+ ret bfloat %result
+}
+
+define void @global_system_atomic_fmin_noret_bf16__offset12b_pos(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_system_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB41_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB41_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB41_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB41_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB41_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB41_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB41_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB41_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v5, v5
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB41_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, bfloat %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x half>
+; --------------------------------------------------------------------
+
+define <2 x half> @global_agent_atomic_fmin_ret_v2f16(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v3, v4, v4
+; GFX12-NEXT: v_pk_min_num_f16 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB42_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX940-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_min_f16 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB42_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX11-NEXT: v_pk_min_f16 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB42_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX10-NEXT: v_pk_min_f16 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB42_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX90A-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX90A-NEXT: v_pk_min_f16 v4, v3, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB42_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX908-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX908-NEXT: v_pk_min_f16 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB42_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v4, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v3
+; GFX8-NEXT: v_max_f16_sdwa v3, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v6, v6
+; GFX8-NEXT: v_min_f16_sdwa v3, v3, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB42_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX7-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX7-NEXT: v_mov_b32_e32 v9, v7
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB42_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_v2f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX6-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_min_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX6-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX6-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX6-NEXT: v_mov_b32_e32 v9, v7
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB42_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v2
+; GFX6-NEXT: v_mov_b32_e32 v1, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(1) %ptr, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define <2 x half> @global_agent_atomic_fmin_ret_v2f16__offset12b_pos(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v3, v4, v4
+; GFX12-NEXT: v_pk_min_num_f16 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB43_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX940-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_min_f16 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB43_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX11-NEXT: v_pk_min_f16 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB43_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX10-NEXT: v_pk_min_f16 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB43_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX90A-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX90A-NEXT: v_pk_min_f16 v4, v3, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB43_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX908-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX908-NEXT: v_pk_min_f16 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB43_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v1, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_max_f16_sdwa v0, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v6, v6
+; GFX8-NEXT: v_min_f16_sdwa v0, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB43_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX7-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX7-NEXT: v_mov_b32_e32 v9, v7
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB43_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX6-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_min_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX6-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX6-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX6-NEXT: v_mov_b32_e32 v9, v7
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB43_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v2
+; GFX6-NEXT: v_mov_b32_e32 v1, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fmin ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define <2 x half> @global_agent_atomic_fmin_ret_v2f16__offset12b_neg(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_v2f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v3, v4, v4
+; GFX12-NEXT: v_pk_min_num_f16 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB44_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_v2f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX940-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_min_f16 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB44_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_v2f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX11-NEXT: v_pk_min_f16 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB44_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_v2f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX10-NEXT: v_pk_min_f16 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB44_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_v2f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX90A-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX90A-NEXT: v_pk_min_f16 v4, v3, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB44_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_v2f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX908-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX908-NEXT: v_pk_min_f16 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB44_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_v2f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v1, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_max_f16_sdwa v0, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v6, v6
+; GFX8-NEXT: v_min_f16_sdwa v0, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB44_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_v2f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v6, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v2
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v0
+; GFX7-NEXT: v_mov_b32_e32 v9, v7
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[8:9], v[4:5], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB44_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_v2f16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v6, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v2
+; GFX6-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v6
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX6-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX6-NEXT: v_or_b32_e32 v6, v8, v0
+; GFX6-NEXT: v_mov_b32_e32 v9, v7
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[8:9], v[4:5], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB44_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 -512
+ %result = atomicrmw fmin ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define void @global_agent_atomic_fmin_noret_v2f16(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_pk_max_num_f16 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_min_num_f16 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB45_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX940-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB45_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB45_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX10-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB45_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX90A-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX90A-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB45_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX908-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX908-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB45_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v4, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v2, v2
+; GFX8-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v2, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v6, v3, v3
+; GFX8-NEXT: v_min_f16_sdwa v2, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v6, v6, v5
+; GFX8-NEXT: v_or_b32_e32 v2, v6, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB45_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB45_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_v2f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX6-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX6-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB45_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmin ptr addrspace(1) %ptr, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmin_noret_v2f16__offset12b_pos(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_pk_max_num_f16 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_min_num_f16 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB46_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX940-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB46_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB46_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX10-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB46_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX90A-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX90A-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB46_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX908-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX908-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB46_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v4, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v2, v2
+; GFX8-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v2, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v6, v3, v3
+; GFX8-NEXT: v_min_f16_sdwa v2, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v6, v6, v5
+; GFX8-NEXT: v_or_b32_e32 v2, v6, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB46_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB46_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX6-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX6-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB46_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmin_noret_v2f16__offset12b_neg(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_v2f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: v_pk_max_num_f16 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_min_num_f16 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB47_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_v2f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX940-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB47_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_v2f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB47_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_v2f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX10-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB47_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_v2f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX90A-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX90A-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB47_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_v2f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX908-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX908-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB47_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_v2f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v4, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v2, v2
+; GFX8-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v2, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v6, v3, v3
+; GFX8-NEXT: v_min_f16_sdwa v2, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v6, v6, v5
+; GFX8-NEXT: v_or_b32_e32 v2, v6, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB47_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_v2f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB47_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_v2f16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX6-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX6-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB47_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 -512
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x half> @global_system_atomic_fmin_ret_v2f16__offset12b_pos(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_system_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v3, v4, v4
+; GFX12-NEXT: v_pk_min_num_f16 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB48_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX940-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_min_f16 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB48_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX11-NEXT: v_pk_min_f16 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB48_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX10-NEXT: v_pk_min_f16 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB48_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX90A-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX90A-NEXT: v_pk_min_f16 v4, v3, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB48_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX908-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX908-NEXT: v_pk_min_f16 v3, v3, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB48_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v1, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_max_f16_sdwa v0, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v6, v6
+; GFX8-NEXT: v_min_f16_sdwa v0, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB48_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX7-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX7-NEXT: v_mov_b32_e32 v9, v7
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB48_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX6-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_min_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX6-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX6-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX6-NEXT: v_mov_b32_e32 v9, v7
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB48_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v2
+; GFX6-NEXT: v_mov_b32_e32 v1, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fmin ptr addrspace(1) %gep, <2 x half> %val seq_cst
+ ret <2 x half> %result
+}
+
+define void @global_system_atomic_fmin_noret_v2f16__offset12b_pos(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_system_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_pk_max_num_f16 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_min_num_f16 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB49_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX940-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB49_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB49_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX10-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB49_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX90A-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX90A-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB49_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX908-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX908-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB49_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v4, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v2, v2
+; GFX8-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v2, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v6, v3, v3
+; GFX8-NEXT: v_min_f16_sdwa v2, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v6, v6, v5
+; GFX8-NEXT: v_or_b32_e32 v2, v6, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB49_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB49_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX6-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX6-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB49_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, <2 x half> %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x bfloat>
+; --------------------------------------------------------------------
+
+define <2 x bfloat> @global_agent_atomic_fmin_ret_v2bf16(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_min_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB50_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB50_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB50_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB50_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB50_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB50_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v3
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v3, 16
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB50_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v2, v2, v7, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v3
+; GFX7-NEXT: v_mov_b32_e32 v6, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB50_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_v2bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX6-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_min_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v2, v2, v7, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v3
+; GFX6-NEXT: v_mov_b32_e32 v6, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB50_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr addrspace(1) %ptr, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define <2 x bfloat> @global_agent_atomic_fmin_ret_v2bf16__offset12b_pos(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_min_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB51_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB51_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB51_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB51_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB51_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB51_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_min_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB51_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v2, v2, v7, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v3
+; GFX7-NEXT: v_mov_b32_e32 v6, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB51_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX6-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_min_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v2, v2, v7, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v3
+; GFX6-NEXT: v_mov_b32_e32 v6, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB51_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fmin ptr addrspace(1) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define <2 x bfloat> @global_agent_atomic_fmin_ret_v2bf16__offset12b_neg(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_ret_v2bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_min_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB52_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_ret_v2bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB52_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_ret_v2bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB52_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_ret_v2bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB52_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_ret_v2bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB52_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_ret_v2bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB52_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_ret_v2bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_min_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB52_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_ret_v2bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v6, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v0
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v0, v0, v7, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v1
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[4:5], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB52_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_ret_v2bf16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v6, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v0
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX6-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v0
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v0, v0, v7, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v1
+; GFX6-NEXT: v_mov_b32_e32 v6, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[4:5], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v1
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB52_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 -512
+ %result = atomicrmw fmin ptr addrspace(1) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @global_agent_atomic_fmin_noret_v2bf16(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_min_num_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB53_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB53_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB53_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB53_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB53_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB53_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB53_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB53_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_v2bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB53_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmin ptr addrspace(1) %ptr, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmin_noret_v2bf16__offset12b_pos(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_min_num_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB54_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB54_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB54_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB54_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB54_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB54_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB54_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB54_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB54_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fmin_noret_v2bf16__offset12b_neg(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fmin_noret_v2bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_min_num_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB55_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fmin_noret_v2bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB55_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fmin_noret_v2bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB55_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fmin_noret_v2bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB55_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fmin_noret_v2bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB55_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fmin_noret_v2bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB55_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fmin_noret_v2bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB55_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fmin_noret_v2bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB55_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fmin_noret_v2bf16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB55_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 -512
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x bfloat> @global_system_atomic_fmin_ret_v2bf16__offset12b_pos(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_system_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_min_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB56_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB56_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB56_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB56_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB56_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB56_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_min_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB56_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v2, v2, v7, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v3
+; GFX7-NEXT: v_mov_b32_e32 v6, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB56_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX6-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_min_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v2, v2, v7, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v3
+; GFX6-NEXT: v_mov_b32_e32 v6, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB56_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fmin ptr addrspace(1) %gep, <2 x bfloat> %val seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @global_system_atomic_fmin_noret_v2bf16__offset12b_pos(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_system_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_min_num_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB57_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB57_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB57_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB57_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB57_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB57_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB57_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB57_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB57_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fmin ptr addrspace(1) %gep, <2 x bfloat> %val seq_cst
+ ret void
+}
+
+attributes #0 = { nounwind "amdgpu-unsafe-fp-atomics"="true" }
+attributes #1 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" "amdgpu-unsafe-fp-atomics"="true" }
diff --git a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fsub.ll b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fsub.ll
new file mode 100644
index 0000000000000..e26619a39bcef
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fsub.ll
@@ -0,0 +1,17517 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 < %s | FileCheck -check-prefix=GFX12 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx940 < %s | FileCheck -check-prefix=GFX940 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx90a < %s | FileCheck -check-prefix=GFX90A %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx908 < %s | FileCheck -check-prefix=GFX908 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tonga < %s | FileCheck -check-prefix=GFX8 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=hawaii < %s | FileCheck -check-prefix=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tahiti < %s | FileCheck -check-prefix=GFX6 %s
+
+; --------------------------------------------------------------------
+; float
+; --------------------------------------------------------------------
+
+define float @global_agent_atomic_fsub_ret_f32(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB0_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB0_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB0_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB0_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB0_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB0_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB0_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB0_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_f32:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB0_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr addrspace(1) %ptr, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @global_agent_atomic_fsub_ret_f32__offset12b_pos(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB1_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB1_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB1_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB1_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB1_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB1_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB1_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB1_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB1_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fsub ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @global_agent_atomic_fsub_ret_f32__offset12b_neg(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB2_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB2_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB2_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB2_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB2_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB2_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB2_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v0, v[3:4], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v3
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, -1, v4, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: v_sub_f32_e32 v5, v6, v2
+; GFX7-NEXT: v_mov_b32_e32 v0, v5
+; GFX7-NEXT: v_mov_b32_e32 v1, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[0:1], v[3:4], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v0, v[3:4], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v3
+; GFX6-NEXT: v_addc_u32_e32 v4, vcc, -1, v4, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v0
+; GFX6-NEXT: v_sub_f32_e32 v5, v6, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v0, v5
+; GFX6-NEXT: v_mov_b32_e32 v1, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], v[3:4], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB2_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 -512
+ %result = atomicrmw fsub ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define void @global_agent_atomic_fsub_noret_f32(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB3_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB3_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB3_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB3_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB3_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB3_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB3_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB3_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_f32:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB3_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fsub ptr addrspace(1) %ptr, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fsub_noret_f32__offset12b_pos(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB4_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v5, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB4_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB4_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB4_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB4_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB4_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB4_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB4_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB4_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fsub_noret_f32__offset12b_neg(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB5_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v5, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB5_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off offset:-2048
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB5_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:-2048
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB5_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB5_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v4, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB5_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 -512
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define float @global_system_atomic_fsub_ret_f32__offset12b_pos(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_system_atomic_fsub_ret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB6_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fsub_ret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB6_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fsub_ret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB6_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fsub_ret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB6_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fsub_ret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB6_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fsub_ret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB6_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fsub_ret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB6_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fsub_ret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB6_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fsub_ret_f32__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB6_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fsub ptr addrspace(1) %gep, float %val seq_cst
+ ret float %result
+}
+
+define void @global_system_atomic_fsub_noret_f32__offset12b_pos(ptr addrspace(1) %ptr, float %val) #0 {
+; GFX12-LABEL: global_system_atomic_fsub_noret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB7_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fsub_noret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v5, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB7_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fsub_noret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB7_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fsub_noret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB7_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fsub_noret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB7_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fsub_noret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB7_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fsub_noret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB7_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fsub_noret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB7_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fsub_noret_f32__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB7_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, float %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; float with ftz/daz
+; --------------------------------------------------------------------
+
+define float @global_agent_atomic_fsub_ret_f32__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_f32__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB8_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_f32__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB8_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_f32__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB8_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_f32__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB8_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_f32__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB8_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_f32__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_f32__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_f32__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB8_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_f32__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB8_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr addrspace(1) %ptr, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @global_agent_atomic_fsub_ret_f32__offset12b_pos__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB9_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB9_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB9_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB9_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB9_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB9_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB9_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB9_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB9_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fsub ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @global_agent_atomic_fsub_ret_f32__offset12b_neg__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_neg__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB10_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_neg__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB10_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_neg__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB10_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_neg__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB10_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_neg__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB10_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_neg__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB10_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_neg__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB10_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_neg__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v0, v[3:4], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v3
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, -1, v4, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: v_sub_f32_e32 v5, v6, v2
+; GFX7-NEXT: v_mov_b32_e32 v0, v5
+; GFX7-NEXT: v_mov_b32_e32 v1, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[0:1], v[3:4], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB10_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_f32__offset12b_neg__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v0, v[3:4], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v3
+; GFX6-NEXT: v_addc_u32_e32 v4, vcc, -1, v4, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v0
+; GFX6-NEXT: v_sub_f32_e32 v5, v6, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v0, v5
+; GFX6-NEXT: v_mov_b32_e32 v1, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], v[3:4], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB10_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 -512
+ %result = atomicrmw fsub ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define void @global_agent_atomic_fsub_noret_f32__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_f32__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB11_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_f32__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB11_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_f32__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB11_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_f32__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB11_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_f32__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_f32__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_f32__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_f32__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB11_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_f32__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB11_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fsub ptr addrspace(1) %ptr, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fsub_noret_f32__offset12b_pos__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB12_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v5, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB12_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB12_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB12_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB12_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB12_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB12_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB12_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB12_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fsub_noret_f32__offset12b_neg__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_neg__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB13_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_neg__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v5, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB13_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_neg__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off offset:-2048
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB13_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_neg__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:-2048
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB13_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_neg__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB13_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_neg__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v4, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB13_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_neg__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB13_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_neg__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB13_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_f32__offset12b_neg__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB13_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 -512
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define float @global_system_atomic_fsub_ret_f32__offset12b_pos__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_system_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB14_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB14_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB14_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB14_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[3:4], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB14_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fsub ptr addrspace(1) %gep, float %val seq_cst
+ ret float %result
+}
+
+define void @global_system_atomic_fsub_noret_f32__offset12b_pos__ftz(ptr addrspace(1) %ptr, float %val) #1 {
+; GFX12-LABEL: global_system_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB15_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v5, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB15_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB15_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB15_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB15_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB15_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB15_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB15_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB15_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, float %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; double
+; --------------------------------------------------------------------
+
+define double @global_agent_atomic_fsub_ret_f64(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[4:5], v[0:1], off
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_add_f64_e64 v[4:5], v[6:7], -v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB16_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB16_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[4:5], v[0:1], off
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB16_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB16_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX90A-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB16_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[4:5], v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB16_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v7, v5
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB16_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: v_mov_b32_e32 v1, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_mov_b32_e32 v7, v1
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v[6:7], s[4:7], 0 addr64
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v11, v1
+; GFX7-NEXT: v_mov_b32_e32 v10, v0
+; GFX7-NEXT: v_add_f64 v[8:9], v[10:11], -v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v8
+; GFX7-NEXT: v_mov_b32_e32 v1, v9
+; GFX7-NEXT: v_mov_b32_e32 v2, v10
+; GFX7-NEXT: v_mov_b32_e32 v3, v11
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[6:7], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB16_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_f64:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_mov_b32_e32 v7, v1
+; GFX6-NEXT: v_mov_b32_e32 v6, v0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dwordx2 v[0:1], v[6:7], s[4:7], 0 addr64
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v1
+; GFX6-NEXT: v_mov_b32_e32 v10, v0
+; GFX6-NEXT: v_add_f64 v[8:9], v[10:11], -v[4:5]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v0, v8
+; GFX6-NEXT: v_mov_b32_e32 v1, v9
+; GFX6-NEXT: v_mov_b32_e32 v2, v10
+; GFX6-NEXT: v_mov_b32_e32 v3, v11
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[6:7], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB16_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr addrspace(1) %ptr, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define double @global_agent_atomic_fsub_ret_f64__offset12b_pos(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_f64__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[4:5], v[0:1], off offset:2040
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_add_f64_e64 v[4:5], v[6:7], -v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:2040 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB17_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_f64__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:2040 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB17_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_f64__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[4:5], v[0:1], off offset:2040
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB17_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_f64__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB17_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_f64__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX90A-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB17_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_f64__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:2040
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_f64__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7f8, v0
+; GFX8-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dwordx2 v[0:1], v[4:5]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v9, v1
+; GFX8-NEXT: v_mov_b32_e32 v8, v0
+; GFX8-NEXT: v_add_f64 v[6:7], v[8:9], -v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_f64__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_mov_b32_e32 v7, v1
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v[6:7], s[4:7], 0 addr64 offset:2040
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v11, v1
+; GFX7-NEXT: v_mov_b32_e32 v10, v0
+; GFX7-NEXT: v_add_f64 v[8:9], v[10:11], -v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v8
+; GFX7-NEXT: v_mov_b32_e32 v1, v9
+; GFX7-NEXT: v_mov_b32_e32 v2, v10
+; GFX7-NEXT: v_mov_b32_e32 v3, v11
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[6:7], s[4:7], 0 addr64 offset:2040 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_f64__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_mov_b32_e32 v7, v1
+; GFX6-NEXT: v_mov_b32_e32 v6, v0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dwordx2 v[0:1], v[6:7], s[4:7], 0 addr64 offset:2040
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v1
+; GFX6-NEXT: v_mov_b32_e32 v10, v0
+; GFX6-NEXT: v_add_f64 v[8:9], v[10:11], -v[4:5]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v0, v8
+; GFX6-NEXT: v_mov_b32_e32 v1, v9
+; GFX6-NEXT: v_mov_b32_e32 v2, v10
+; GFX6-NEXT: v_mov_b32_e32 v3, v11
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[6:7], s[4:7], 0 addr64 offset:2040 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB17_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(1) %ptr, i64 255
+ %result = atomicrmw fsub ptr addrspace(1) %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define double @global_agent_atomic_fsub_ret_f64__offset12b_neg(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_f64__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[4:5], v[0:1], off offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_add_f64_e64 v[4:5], v[6:7], -v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB18_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_f64__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB18_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_f64__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[4:5], v[0:1], off offset:-2048
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB18_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_f64__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB18_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_f64__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX90A-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB18_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_f64__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[4:5], v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB18_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_f64__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dwordx2 v[0:1], v[4:5]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v9, v1
+; GFX8-NEXT: v_mov_b32_e32 v8, v0
+; GFX8-NEXT: v_add_f64 v[6:7], v[8:9], -v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB18_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_f64__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: v_mov_b32_e32 v7, v1
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v[6:7], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v6, vcc, 0xfffff800, v6
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: v_addc_u32_e32 v7, vcc, -1, v7, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v11, v1
+; GFX7-NEXT: v_mov_b32_e32 v10, v0
+; GFX7-NEXT: v_add_f64 v[8:9], v[10:11], -v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v8
+; GFX7-NEXT: v_mov_b32_e32 v1, v9
+; GFX7-NEXT: v_mov_b32_e32 v2, v10
+; GFX7-NEXT: v_mov_b32_e32 v3, v11
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[6:7], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB18_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_f64__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: v_mov_b32_e32 v7, v1
+; GFX6-NEXT: v_mov_b32_e32 v6, v0
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dwordx2 v[0:1], v[6:7], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v6, vcc, 0xfffff800, v6
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: v_addc_u32_e32 v7, vcc, -1, v7, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v1
+; GFX6-NEXT: v_mov_b32_e32 v10, v0
+; GFX6-NEXT: v_add_f64 v[8:9], v[10:11], -v[4:5]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v0, v8
+; GFX6-NEXT: v_mov_b32_e32 v1, v9
+; GFX6-NEXT: v_mov_b32_e32 v2, v10
+; GFX6-NEXT: v_mov_b32_e32 v3, v11
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v[6:7], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB18_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(1) %ptr, i64 -256
+ %result = atomicrmw fsub ptr addrspace(1) %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define void @global_agent_atomic_fsub_noret_f64(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[6:7], v[0:1], off
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_add_f64_e64 v[4:5], v[6:7], -v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB19_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dwordx2 v[6:7], v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB19_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[6:7], v[0:1], off
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB19_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[6:7], v[0:1], off
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB19_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx2 v[6:7], v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX90A-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB19_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[6:7], v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB19_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dwordx2 v[6:7], v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v7, v5
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB19_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dwordx2 v[6:7], v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX7-NEXT: v_mov_b32_e32 v11, v7
+; GFX7-NEXT: v_mov_b32_e32 v10, v6
+; GFX7-NEXT: v_mov_b32_e32 v9, v5
+; GFX7-NEXT: v_mov_b32_e32 v8, v4
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v6, v8
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v7, v9
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB19_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_f64:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dwordx2 v[6:7], v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v7
+; GFX6-NEXT: v_mov_b32_e32 v10, v6
+; GFX6-NEXT: v_mov_b32_e32 v9, v5
+; GFX6-NEXT: v_mov_b32_e32 v8, v4
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v6, v8
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v7, v9
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB19_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fsub ptr addrspace(1) %ptr, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fsub_noret_f64__offset12b_pos(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_f64__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[6:7], v[0:1], off offset:2040
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_add_f64_e64 v[4:5], v[6:7], -v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:2040 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB20_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_f64__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dwordx2 v[6:7], v[0:1], off offset:2040
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:2040 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB20_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_f64__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[6:7], v[0:1], off offset:2040
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB20_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_f64__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[6:7], v[0:1], off offset:2040
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB20_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_f64__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx2 v[6:7], v[0:1], off offset:2040
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX90A-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB20_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_f64__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[6:7], v[0:1], off offset:2040
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:2040 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB20_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_f64__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7f8, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dwordx2 v[6:7], v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v7, v5
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB20_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_f64__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dwordx2 v[6:7], v[0:1], s[4:7], 0 addr64 offset:2040
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX7-NEXT: v_mov_b32_e32 v11, v7
+; GFX7-NEXT: v_mov_b32_e32 v10, v6
+; GFX7-NEXT: v_mov_b32_e32 v9, v5
+; GFX7-NEXT: v_mov_b32_e32 v8, v4
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 offset:2040 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v6, v8
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v7, v9
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB20_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_f64__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dwordx2 v[6:7], v[0:1], s[4:7], 0 addr64 offset:2040
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v7
+; GFX6-NEXT: v_mov_b32_e32 v10, v6
+; GFX6-NEXT: v_mov_b32_e32 v9, v5
+; GFX6-NEXT: v_mov_b32_e32 v8, v4
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 offset:2040 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v6, v8
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v7, v9
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB20_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(1) %ptr, i64 255
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fsub_noret_f64__offset12b_neg(ptr addrspace(1) %ptr, double %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_f64__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b64 v[6:7], v[0:1], off offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_add_f64_e64 v[4:5], v[6:7], -v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB21_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_f64__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dwordx2 v[6:7], v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB21_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_f64__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b64 v[6:7], v[0:1], off offset:-2048
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB21_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_f64__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dwordx2 v[6:7], v[0:1], off offset:-2048
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB21_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_f64__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx2 v[6:7], v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX90A-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB21_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_f64__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dwordx2 v[6:7], v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX908-NEXT: global_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB21_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_f64__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dwordx2 v[6:7], v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v7, v5
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB21_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_f64__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dwordx2 v[6:7], v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX7-NEXT: v_mov_b32_e32 v11, v7
+; GFX7-NEXT: v_mov_b32_e32 v10, v6
+; GFX7-NEXT: v_mov_b32_e32 v9, v5
+; GFX7-NEXT: v_mov_b32_e32 v8, v4
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v6, v8
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v7, v9
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB21_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_f64__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dwordx2 v[6:7], v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v11, v7
+; GFX6-NEXT: v_mov_b32_e32 v10, v6
+; GFX6-NEXT: v_mov_b32_e32 v9, v5
+; GFX6-NEXT: v_mov_b32_e32 v8, v4
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v6, v8
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v7, v9
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB21_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(1) %ptr, i64 -256
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; half
+; --------------------------------------------------------------------
+
+define half @global_agent_atomic_fsub_ret_f16(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB22_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: global_load_dword v4, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX940-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB22_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB22_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB22_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v4, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX90A-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB22_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v4
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX908-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX908-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX908-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB22_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB22_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v5
+; GFX7-NEXT: v_not_b32_e32 v7, v2
+; GFX7-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v5, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v3, v4, v7
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v5, v2
+; GFX7-NEXT: v_or_b32_e32 v3, v3, v2
+; GFX7-NEXT: v_mov_b32_e32 v2, v3
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[2:3], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB22_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v5, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v5
+; GFX6-NEXT: v_not_b32_e32 v7, v2
+; GFX6-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v5, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v3, v4, v7
+; GFX6-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v5, v2
+; GFX6-NEXT: v_or_b32_e32 v3, v3, v2
+; GFX6-NEXT: v_mov_b32_e32 v2, v3
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[2:3], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB22_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v5, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr addrspace(1) %ptr, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+define half @global_agent_atomic_fsub_ret_f16__offset12b_pos(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB23_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[6:7], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v6
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_load_dword v4, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX940-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB23_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB23_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB23_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v4, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX90A-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB23_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v4
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX908-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX908-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX908-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB23_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB23_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX7-NEXT: v_not_b32_e32 v8, v2
+; GFX7-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v7
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB23_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX6-NEXT: v_not_b32_e32 v8, v2
+; GFX6-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX6-NEXT: v_sub_f32_e32 v2, v2, v7
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB23_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fsub ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+define half @global_agent_atomic_fsub_ret_f16__offset12b_neg(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB24_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[6:7], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v6
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_load_dword v4, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX940-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB24_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB24_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB24_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v4, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX90A-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB24_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v4
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX908-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX908-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX908-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB24_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB24_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX7-NEXT: v_not_b32_e32 v8, v2
+; GFX7-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v7
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB24_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX6-NEXT: v_not_b32_e32 v8, v2
+; GFX6-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX6-NEXT: v_sub_f32_e32 v2, v2, v7
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB24_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 -1024
+ %result = atomicrmw fsub ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+ }
+
+define void @global_agent_atomic_fsub_noret_f16(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB25_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX940-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB25_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB25_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v4, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX10-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB25_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX90A-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB25_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX908-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB25_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX8-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB25_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: v_lshl_b32_e32 v3, 0xffff, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_not_b32_e32 v6, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v2, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX7-NEXT: v_sub_f32_e32 v3, v3, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v2, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX7-NEXT: v_mov_b32_e32 v8, v4
+; GFX7-NEXT: v_mov_b32_e32 v7, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB25_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v2
+; GFX6-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX6-NEXT: v_lshl_b32_e32 v3, 0xffff, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_not_b32_e32 v6, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v2, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX6-NEXT: v_sub_f32_e32 v3, v3, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v2, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX6-NEXT: v_mov_b32_e32 v8, v4
+; GFX6-NEXT: v_mov_b32_e32 v7, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB25_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fsub ptr addrspace(1) %ptr, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fsub_noret_f16__offset12b_pos(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB26_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX940-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB26_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB26_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v4, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX10-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB26_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX90A-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB26_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX908-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB26_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX8-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB26_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB26_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX6-NEXT: v_not_b32_e32 v6, v2
+; GFX6-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX6-NEXT: v_sub_f32_e32 v2, v2, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB26_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fsub_noret_f16__offset12b_neg(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB27_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX940-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB27_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB27_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v4, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX10-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB27_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX90A-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB27_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX908-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB27_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX8-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB27_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB27_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_f16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX6-NEXT: v_not_b32_e32 v6, v2
+; GFX6-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX6-NEXT: v_sub_f32_e32 v2, v2, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB27_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 -1024
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define half @global_agent_atomic_fsub_ret_f16__offset12b_pos__align4(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_pos__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f16_e32 v3, v4, v2
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB28_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_pos__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_sub_f16_e32 v3, v5, v2
+; GFX940-NEXT: v_and_or_b32 v4, v5, s2, v3
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB28_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_pos__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f16_e32 v3, v4, v2
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB28_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_pos__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_sub_f16_e32 v3, v4, v2
+; GFX10-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX10-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB28_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_pos__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_sub_f16_e32 v3, v5, v2
+; GFX90A-NEXT: v_and_or_b32 v4, v5, s6, v3
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB28_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_pos__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_sub_f16_e32 v3, v4, v2
+; GFX908-NEXT: v_and_or_b32 v3, v4, s6, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB28_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_pos__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_sub_f16_e32 v0, v1, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX8-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB28_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_pos__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX7-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB28_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v4
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_f16__offset12b_pos__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v2
+; GFX6-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX6-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB28_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fsub ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst, align 4
+ ret half %result
+}
+
+define void @global_agent_atomic_fsub_noret_f16__offset12b__align4_pos(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_f16__offset12b__align4_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off offset:2046
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_sub_f16_e32 v3, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB29_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_f16__offset12b__align4_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v5, v[0:1], off offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_sub_f16_e32 v3, v5, v2
+; GFX940-NEXT: v_and_or_b32 v4, v5, s2, v3
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB29_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_f16__offset12b__align4_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off offset:2046
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_sub_f16_e32 v3, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB29_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_f16__offset12b__align4_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:2046
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_sub_f16_e32 v3, v4, v2
+; GFX10-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX10-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB29_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_f16__offset12b__align4_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_sub_f16_e32 v3, v5, v2
+; GFX90A-NEXT: v_and_or_b32 v4, v5, s6, v3
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB29_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_f16__offset12b__align4_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v4, v[0:1], off offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_sub_f16_e32 v3, v4, v2
+; GFX908-NEXT: v_and_or_b32 v3, v4, s6, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB29_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_f16__offset12b__align4_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_sub_f16_e32 v3, v4, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB29_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_f16__offset12b__align4_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v2
+; GFX7-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB29_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_f16__offset12b__align4_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v2
+; GFX6-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX6-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB29_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, half %val syncscope("agent") seq_cst, align 4
+ ret void
+}
+
+define half @global_system_atomic_fsub_ret_f16__offset12b_pos(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_system_atomic_fsub_ret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB30_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fsub_ret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[6:7], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v6
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_load_dword v4, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX940-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB30_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fsub_ret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB30_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fsub_ret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB30_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fsub_ret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v4, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX90A-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB30_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fsub_ret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v4
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX908-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX908-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX908-NEXT: global_atomic_cmpswap v4, v[0:1], v[6:7], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB30_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fsub_ret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB30_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fsub_ret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX7-NEXT: v_not_b32_e32 v8, v2
+; GFX7-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v7
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB30_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fsub_ret_f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v6
+; GFX6-NEXT: v_not_b32_e32 v8, v2
+; GFX6-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v8
+; GFX6-NEXT: v_sub_f32_e32 v2, v2, v7
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB30_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fsub ptr addrspace(1) %gep, half %val seq_cst
+ ret half %result
+}
+
+define void @global_system_atomic_fsub_noret_f16__offset12b_pos(ptr addrspace(1) %ptr, half %val) #0 {
+; GFX12-LABEL: global_system_atomic_fsub_noret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB31_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fsub_noret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX940-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB31_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fsub_noret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB31_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fsub_noret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v4, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX10-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB31_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fsub_noret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX90A-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB31_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fsub_noret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX908-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB31_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fsub_noret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX8-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB31_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fsub_noret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB31_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fsub_noret_f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX6-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX6-NEXT: v_not_b32_e32 v6, v2
+; GFX6-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX6-NEXT: v_sub_f32_e32 v2, v2, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB31_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, half %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; bfloat
+; --------------------------------------------------------------------
+
+define bfloat @global_agent_atomic_fsub_ret_bf16(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB32_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB32_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB32_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB32_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB32_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB32_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB32_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v3, 0xffff, v5
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v6, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX7-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v5, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v7
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v3, v4, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v5, v2
+; GFX7-NEXT: v_or_b32_e32 v3, v3, v2
+; GFX7-NEXT: v_mov_b32_e32 v2, v3
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[2:3], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB32_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v5, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX6-NEXT: v_lshl_b32_e32 v3, 0xffff, v5
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v6, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX6-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v5, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_sub_f32_e32 v2, v2, v7
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_and_b32_e32 v3, v4, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v5, v2
+; GFX6-NEXT: v_or_b32_e32 v3, v3, v2
+; GFX6-NEXT: v_mov_b32_e32 v2, v3
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[2:3], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB32_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v5, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr addrspace(1) %ptr, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+define bfloat @global_agent_atomic_fsub_ret_bf16__offset12b_pos(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB33_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB33_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB33_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB33_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB33_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB33_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB33_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v7, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX7-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB33_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v7, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX6-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_sub_f32_e32 v2, v2, v8
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB33_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fsub ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+define bfloat @global_agent_atomic_fsub_ret_bf16__offset12b_neg(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB34_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB34_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB34_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB34_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB34_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB34_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB34_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v7, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX7-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB34_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v7, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX6-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_sub_f32_e32 v2, v2, v8
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB34_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 -1024
+ %result = atomicrmw fsub ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+ }
+
+define void @global_agent_atomic_fsub_noret_bf16(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB35_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v4, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_sub_f32_e32 v4, v4, v2
+; GFX940-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX940-NEXT: v_add3_u32 v7, v7, v4, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v4, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB35_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB35_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: global_load_dword v4, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB35_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_sub_f32_e32 v4, v4, v2
+; GFX90A-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX90A-NEXT: v_add3_u32 v7, v7, v4, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: global_atomic_cmpswap v4, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB35_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB35_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX8-NEXT: v_bfe_u32 v8, v3, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v3
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB35_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v3, 0xffff, v5
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v6, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX7-NEXT: v_mov_b32_e32 v8, v4
+; GFX7-NEXT: v_mov_b32_e32 v7, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v4
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v4, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB35_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v3, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v4, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX6-NEXT: v_lshl_b32_e32 v3, 0xffff, v5
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v6, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX6-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX6-NEXT: v_mov_b32_e32 v8, v4
+; GFX6-NEXT: v_mov_b32_e32 v7, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v4
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v4, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB35_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fsub ptr addrspace(1) %ptr, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fsub_noret_bf16__offset12b_pos(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB36_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB36_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB36_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB36_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB36_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB36_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB36_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB36_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v5, v5
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB36_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fsub_noret_bf16__offset12b_neg(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB37_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB37_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB37_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB37_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB37_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB37_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB37_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB37_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v5, v5
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB37_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 -1024
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define bfloat @global_agent_atomic_fsub_ret_bf16__offset12b_pos__align4(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_pos__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX12-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB38_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_pos__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX940-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX940-NEXT: v_bfe_u32 v4, v3, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX940-NEXT: v_add3_u32 v4, v4, v3, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v3, v4, v6, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX940-NEXT: v_and_or_b32 v4, v5, s3, v3
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB38_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_pos__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX11-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB38_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_pos__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX10-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX10-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX10-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX10-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB38_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_pos__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX90A-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX90A-NEXT: v_bfe_u32 v4, v3, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX90A-NEXT: v_add3_u32 v4, v4, v3, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e32 v3, v4, v6, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX90A-NEXT: v_and_or_b32 v4, v5, s7, v3
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB38_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_pos__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX908-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX908-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, s7, v3
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB38_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_pos__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_sub_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v0
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX8-NEXT: v_cndmask_b32_e32 v0, v5, v7, vcc
+; GFX8-NEXT: v_or_b32_sdwa v5, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB38_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_pos__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB38_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v4
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_bf16__offset12b_pos__align4:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX6-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB38_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fsub ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst, align 4
+ ret bfloat %result
+}
+
+define void @global_agent_atomic_fsub_noret_bf16__offset12b__align4_pos(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b__align4_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB39_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b__align4_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v5, v5, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: v_and_or_b32 v2, v3, s3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB39_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b__align4_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2046
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB39_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b__align4_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB39_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b__align4_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v5, v5, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v3, s7, v2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB39_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b__align4_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v5, v5, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: v_and_or_b32 v2, v3, s7, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB39_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b__align4_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v2
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v6, v7, vcc
+; GFX8-NEXT: v_or_b32_sdwa v2, v5, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB39_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b__align4_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX7-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB39_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b__align4_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64 offset:2046
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX6-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX6-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v[0:1], s[4:7], 0 addr64 offset:2046 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB39_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, bfloat %val syncscope("agent") seq_cst, align 4
+ ret void
+}
+
+define bfloat @global_system_atomic_fsub_ret_bf16__offset12b_pos(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_system_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v5, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB40_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB40_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v5, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB40_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v5, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB40_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v5, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB40_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: global_load_dword v5, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: global_atomic_cmpswap v5, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB40_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB40_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v7, v4
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX7-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB40_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v6, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v4, 0xffff, v6
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v7, v4
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v8, 0xffff0000, v2
+; GFX6-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_sub_f32_e32 v2, v2, v8
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, v3, v7
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v6, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: v_mov_b32_e32 v4, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB40_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, v6, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %result = atomicrmw fsub ptr addrspace(1) %gep, bfloat %val seq_cst
+ ret bfloat %result
+}
+
+define void @global_system_atomic_fsub_noret_bf16__offset12b_pos(ptr addrspace(1) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: global_system_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB41_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB41_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB41_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB41_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB41_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB41_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB41_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v7
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB41_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX6-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_not_b32_e32 v5, v5
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX6-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX6-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v3
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v7
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB41_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(1) %ptr, i64 1023
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, bfloat %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x half>
+; --------------------------------------------------------------------
+
+define <2 x half> @global_agent_atomic_fsub_ret_v2f16(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB42_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB42_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB42_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB42_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB42_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB42_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_sub_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_sub_f16_e32 v5, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB42_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX7-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX7-NEXT: v_mov_b32_e32 v9, v7
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB42_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_v2f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX6-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_sub_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX6-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX6-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX6-NEXT: v_mov_b32_e32 v9, v7
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB42_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v2
+; GFX6-NEXT: v_mov_b32_e32 v1, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr addrspace(1) %ptr, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define <2 x half> @global_agent_atomic_fsub_ret_v2f16__offset12b_pos(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB43_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB43_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB43_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB43_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB43_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB43_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_sub_f16_sdwa v0, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_sub_f16_e32 v5, v1, v2
+; GFX8-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB43_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX7-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX7-NEXT: v_mov_b32_e32 v9, v7
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB43_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX6-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_sub_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX6-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX6-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX6-NEXT: v_mov_b32_e32 v9, v7
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB43_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v2
+; GFX6-NEXT: v_mov_b32_e32 v1, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fsub ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define <2 x half> @global_agent_atomic_fsub_ret_v2f16__offset12b_neg(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_v2f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB44_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_v2f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB44_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_v2f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB44_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_v2f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB44_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_v2f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB44_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_v2f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB44_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_v2f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_sub_f16_sdwa v0, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_sub_f16_e32 v5, v1, v2
+; GFX8-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB44_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_v2f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v6, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v2
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v0
+; GFX7-NEXT: v_mov_b32_e32 v9, v7
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[8:9], v[4:5], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB44_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_v2f16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v6, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v2
+; GFX6-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v6
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX6-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX6-NEXT: v_or_b32_e32 v6, v8, v0
+; GFX6-NEXT: v_mov_b32_e32 v9, v7
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[8:9], v[4:5], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB44_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 -512
+ %result = atomicrmw fsub ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define void @global_agent_atomic_fsub_noret_v2f16(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB45_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v5, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB45_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB45_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB45_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB45_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v4, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB45_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_sub_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_sub_f16_e32 v5, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB45_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB45_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_v2f16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX6-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX6-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB45_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fsub ptr addrspace(1) %ptr, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fsub_noret_v2f16__offset12b_pos(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB46_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v5, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB46_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB46_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB46_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB46_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB46_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_sub_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_sub_f16_e32 v5, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB46_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB46_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX6-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX6-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB46_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fsub_noret_v2f16__offset12b_neg(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_v2f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB47_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_v2f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v5, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB47_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_v2f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off offset:-2048
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB47_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_v2f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:-2048
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB47_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_v2f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB47_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_v2f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v4, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB47_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_v2f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_sub_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_sub_f16_e32 v5, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB47_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_v2f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB47_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_v2f16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX6-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX6-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB47_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 -512
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x half> @global_system_atomic_fsub_ret_v2f16__offset12b_pos(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_system_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB48_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB48_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB48_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB48_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB48_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB48_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_sub_f16_sdwa v0, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_sub_f16_e32 v5, v1, v2
+; GFX8-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB48_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX7-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX7-NEXT: v_mov_b32_e32 v9, v7
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB48_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX6-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_sub_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX6-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX6-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX6-NEXT: v_mov_b32_e32 v9, v7
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: buffer_atomic_cmpswap v[8:9], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v8
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB48_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v2
+; GFX6-NEXT: v_mov_b32_e32 v1, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fsub ptr addrspace(1) %gep, <2 x half> %val seq_cst
+ ret <2 x half> %result
+}
+
+define void @global_system_atomic_fsub_noret_v2f16__offset12b_pos(ptr addrspace(1) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: global_system_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v4, v[0:1], off offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB49_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v5, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB49_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v4, v[0:1], off offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB49_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB49_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v5, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB49_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v4, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB49_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_sub_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_sub_f16_e32 v5, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB49_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB49_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX6-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX6-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB49_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, <2 x half> %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x bfloat>
+; --------------------------------------------------------------------
+
+define <2 x bfloat> @global_agent_atomic_fsub_ret_v2bf16(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB50_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB50_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB50_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX10-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB50_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB50_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB50_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX8-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v3
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v3, 16
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB50_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v2, v2, v6, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v3
+; GFX7-NEXT: v_mov_b32_e32 v6, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB50_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_v2bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX6-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_sub_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v7
+; GFX6-NEXT: v_alignbit_b32 v2, v2, v6, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v3
+; GFX6-NEXT: v_mov_b32_e32 v6, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB50_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr addrspace(1) %ptr, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define <2 x bfloat> @global_agent_atomic_fsub_ret_v2bf16__offset12b_pos(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB51_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB51_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB51_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX10-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB51_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB51_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB51_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_sub_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB51_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v2, v2, v6, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v3
+; GFX7-NEXT: v_mov_b32_e32 v6, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB51_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX6-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_sub_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v7
+; GFX6-NEXT: v_alignbit_b32 v2, v2, v6, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v3
+; GFX6-NEXT: v_mov_b32_e32 v6, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB51_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fsub ptr addrspace(1) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define <2 x bfloat> @global_agent_atomic_fsub_ret_v2bf16__offset12b_neg(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_ret_v2bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB52_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_ret_v2bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB52_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_ret_v2bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB52_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_ret_v2bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX10-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB52_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_ret_v2bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB52_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_ret_v2bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB52_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_ret_v2bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_sub_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB52_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_ret_v2bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v6, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v0
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v0, v0, v6, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v1
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[4:5], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB52_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_ret_v2bf16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v6, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v0
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX6-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v1
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v0
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 16, v7
+; GFX6-NEXT: v_alignbit_b32 v0, v0, v6, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v1
+; GFX6-NEXT: v_mov_b32_e32 v6, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[4:5], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v1
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB52_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 -512
+ %result = atomicrmw fsub ptr addrspace(1) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @global_agent_atomic_fsub_noret_v2bf16(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB53_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB53_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB53_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB53_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB53_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB53_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB53_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB53_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_v2bf16:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB53_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fsub ptr addrspace(1) %ptr, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fsub_noret_v2bf16__offset12b_pos(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB54_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB54_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB54_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB54_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB54_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB54_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB54_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB54_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB54_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @global_agent_atomic_fsub_noret_v2bf16__offset12b_neg(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_agent_atomic_fsub_noret_v2bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB55_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_agent_atomic_fsub_noret_v2bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB55_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_agent_atomic_fsub_noret_v2bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-2048
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB55_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_agent_atomic_fsub_noret_v2bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB55_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_agent_atomic_fsub_noret_v2bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB55_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_agent_atomic_fsub_noret_v2bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-2048
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-2048 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB55_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_agent_atomic_fsub_noret_v2bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB55_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_agent_atomic_fsub_noret_v2bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_movk_i32 s4, 0xf800
+; GFX7-NEXT: s_mov_b32 s5, -1
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB55_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_agent_atomic_fsub_noret_v2bf16__offset12b_neg:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_movk_i32 s4, 0xf800
+; GFX6-NEXT: s_mov_b32 s5, -1
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64
+; GFX6-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB55_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 -512
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x bfloat> @global_system_atomic_fsub_ret_v2bf16__offset12b_pos(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_system_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB56_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB56_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB56_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX10-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB56_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB56_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB56_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_sub_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB56_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v2, v2, v6, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v3
+; GFX7-NEXT: v_mov_b32_e32 v6, v2
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB56_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX6-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX6-NEXT: v_sub_f32_e32 v7, v7, v5
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v4
+; GFX6-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v7
+; GFX6-NEXT: v_alignbit_b32 v2, v2, v6, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v3
+; GFX6-NEXT: v_mov_b32_e32 v6, v2
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB56_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v0, v3
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 511
+ %result = atomicrmw fsub ptr addrspace(1) %gep, <2 x bfloat> %val seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @global_system_atomic_fsub_noret_v2bf16__offset12b_pos(ptr addrspace(1) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: global_system_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB57_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: global_system_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB57_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: global_system_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB57_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: global_system_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB57_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: global_system_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB57_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: global_system_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB57_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: global_system_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB57_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: global_system_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 s6, 0
+; GFX7-NEXT: s_mov_b32 s7, 0xf000
+; GFX7-NEXT: s_mov_b32 s4, s6
+; GFX7-NEXT: s_mov_b32 s5, s6
+; GFX7-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB57_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: global_system_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_mov_b32 s6, 0
+; GFX6-NEXT: s_mov_b32 s7, 0xf000
+; GFX6-NEXT: s_mov_b32 s4, s6
+; GFX6-NEXT: s_mov_b32 s5, s6
+; GFX6-NEXT: buffer_load_dword v5, v[0:1], s[4:7], 0 addr64 offset:2044
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX6-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX6-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v[0:1], s[4:7], 0 addr64 offset:2044 glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB57_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i64 511
+ %unused = atomicrmw fsub ptr addrspace(1) %gep, <2 x bfloat> %val seq_cst
+ ret void
+}
+
+attributes #0 = { nounwind "amdgpu-unsafe-fp-atomics"="true" }
+attributes #1 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" "amdgpu-unsafe-fp-atomics"="true" }
diff --git a/llvm/test/CodeGen/AMDGPU/global-atomics-fp.ll b/llvm/test/CodeGen/AMDGPU/global-atomics-fp.ll
deleted file mode 100644
index 88ce07c9779d0..0000000000000
--- a/llvm/test/CodeGen/AMDGPU/global-atomics-fp.ll
+++ /dev/null
@@ -1,8090 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX900 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx908 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX908 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx90a -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX90A %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx940 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX940 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX10 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -amdgpu-enable-delay-alu=0 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX11 %s
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -amdgpu-enable-delay-alu=0 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX12 %s
-
-define amdgpu_kernel void @global_atomic_fadd_ret_f32(ptr addrspace(1) %ptr) #0 {
-; GFX900-LABEL: global_atomic_fadd_ret_f32:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_mov_b64 s[4:5], exec
-; GFX900-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX900-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX900-NEXT: ; implicit-def: $vgpr1
-; GFX900-NEXT: s_and_saveexec_b64 s[2:3], vcc
-; GFX900-NEXT: s_cbranch_execz .LBB0_4
-; GFX900-NEXT: ; %bb.1:
-; GFX900-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX900-NEXT: s_bcnt1_i32_b64 s7, s[4:5]
-; GFX900-NEXT: v_cvt_f32_ubyte0_e32 v1, s7
-; GFX900-NEXT: s_mov_b64 s[4:5], 0
-; GFX900-NEXT: v_mul_f32_e32 v2, 4.0, v1
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: s_load_dword s6, s[0:1], 0x0
-; GFX900-NEXT: v_mov_b32_e32 v3, 0
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v1, s6
-; GFX900-NEXT: .LBB0_2: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: v_mov_b32_e32 v5, v1
-; GFX900-NEXT: v_add_f32_e32 v4, v5, v2
-; GFX900-NEXT: global_atomic_cmpswap v1, v3, v[4:5], s[0:1] glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v1, v5
-; GFX900-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_cbranch_execnz .LBB0_2
-; GFX900-NEXT: ; %bb.3: ; %Flow
-; GFX900-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX900-NEXT: .LBB0_4: ; %Flow1
-; GFX900-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX900-NEXT: v_readfirstlane_b32 s0, v1
-; GFX900-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX900-NEXT: v_mad_f32 v0, v0, 4.0, s0
-; GFX900-NEXT: global_store_dword v[0:1], v0, off
-; GFX900-NEXT: s_endpgm
-;
-; GFX908-LABEL: global_atomic_fadd_ret_f32:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_mov_b64 s[4:5], exec
-; GFX908-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX908-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX908-NEXT: ; implicit-def: $vgpr1
-; GFX908-NEXT: s_and_saveexec_b64 s[2:3], vcc
-; GFX908-NEXT: s_cbranch_execz .LBB0_4
-; GFX908-NEXT: ; %bb.1:
-; GFX908-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX908-NEXT: s_bcnt1_i32_b64 s7, s[4:5]
-; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v1, s7
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: v_mul_f32_e32 v2, 4.0, v1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: s_load_dword s6, s[0:1], 0x0
-; GFX908-NEXT: v_mov_b32_e32 v3, 0
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v1, s6
-; GFX908-NEXT: .LBB0_2: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: v_mov_b32_e32 v5, v1
-; GFX908-NEXT: v_add_f32_e32 v4, v5, v2
-; GFX908-NEXT: global_atomic_cmpswap v1, v3, v[4:5], s[0:1] glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v5
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB0_2
-; GFX908-NEXT: ; %bb.3: ; %Flow
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: .LBB0_4: ; %Flow1
-; GFX908-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX908-NEXT: v_readfirstlane_b32 s0, v1
-; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX908-NEXT: v_mad_f32 v0, v0, 4.0, s0
-; GFX908-NEXT: global_store_dword v[0:1], v0, off
-; GFX908-NEXT: s_endpgm
-;
-; GFX90A-LABEL: global_atomic_fadd_ret_f32:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_mov_b64 s[4:5], exec
-; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX90A-NEXT: ; implicit-def: $vgpr1
-; GFX90A-NEXT: s_and_saveexec_b64 s[2:3], vcc
-; GFX90A-NEXT: s_cbranch_execz .LBB0_4
-; GFX90A-NEXT: ; %bb.1:
-; GFX90A-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX90A-NEXT: s_bcnt1_i32_b64 s7, s[4:5]
-; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v1, s7
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: v_mul_f32_e32 v2, 4.0, v1
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: s_load_dword s6, s[0:1], 0x0
-; GFX90A-NEXT: v_mov_b32_e32 v3, 0
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v1, s6
-; GFX90A-NEXT: .LBB0_2: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: v_mov_b32_e32 v5, v1
-; GFX90A-NEXT: v_add_f32_e32 v4, v5, v2
-; GFX90A-NEXT: buffer_wbl2
-; GFX90A-NEXT: global_atomic_cmpswap v1, v3, v[4:5], s[0:1] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_invl2
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v5
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB0_2
-; GFX90A-NEXT: ; %bb.3: ; %Flow
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: .LBB0_4: ; %Flow1
-; GFX90A-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX90A-NEXT: v_readfirstlane_b32 s0, v1
-; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX90A-NEXT: v_mad_f32 v0, v0, 4.0, s0
-; GFX90A-NEXT: global_store_dword v[0:1], v0, off
-; GFX90A-NEXT: s_endpgm
-;
-; GFX940-LABEL: global_atomic_fadd_ret_f32:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_mov_b64 s[4:5], exec
-; GFX940-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX940-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX940-NEXT: ; implicit-def: $vgpr1
-; GFX940-NEXT: s_and_saveexec_b64 s[2:3], vcc
-; GFX940-NEXT: s_cbranch_execz .LBB0_2
-; GFX940-NEXT: ; %bb.1:
-; GFX940-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX940-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
-; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v2, s4
-; GFX940-NEXT: v_mov_b32_e32 v1, 0
-; GFX940-NEXT: v_mul_f32_e32 v2, 4.0, v2
-; GFX940-NEXT: buffer_wbl2 sc0 sc1
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: global_atomic_add_f32 v1, v1, v2, s[0:1] sc0 sc1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc0 sc1
-; GFX940-NEXT: .LBB0_2:
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX940-NEXT: v_readfirstlane_b32 s0, v1
-; GFX940-NEXT: v_mul_f32_e32 v0, 4.0, v0
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_add_f32_e32 v0, s0, v0
-; GFX940-NEXT: global_store_dword v[0:1], v0, off sc0 sc1
-; GFX940-NEXT: s_endpgm
-;
-; GFX10-LABEL: global_atomic_fadd_ret_f32:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_mov_b32 s4, exec_lo
-; GFX10-NEXT: s_mov_b32 s3, 0
-; GFX10-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX10-NEXT: ; implicit-def: $vgpr1
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
-; GFX10-NEXT: s_and_saveexec_b32 s2, vcc_lo
-; GFX10-NEXT: s_cbranch_execz .LBB0_4
-; GFX10-NEXT: ; %bb.1:
-; GFX10-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX10-NEXT: s_bcnt1_i32_b32 s4, s4
-; GFX10-NEXT: v_mov_b32_e32 v3, 0
-; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
-; GFX10-NEXT: v_mul_f32_e32 v2, 4.0, v1
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: s_load_dword s5, s[0:1], 0x0
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v1, s5
-; GFX10-NEXT: .LBB0_2: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: v_mov_b32_e32 v5, v1
-; GFX10-NEXT: v_add_f32_e32 v4, v5, v2
-; GFX10-NEXT: global_atomic_cmpswap v1, v3, v[4:5], s[0:1] glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v5
-; GFX10-NEXT: s_or_b32 s3, vcc_lo, s3
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s3
-; GFX10-NEXT: s_cbranch_execnz .LBB0_2
-; GFX10-NEXT: ; %bb.3: ; %Flow
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s3
-; GFX10-NEXT: .LBB0_4: ; %Flow1
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s2
-; GFX10-NEXT: v_readfirstlane_b32 s0, v1
-; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX10-NEXT: v_mad_f32 v0, v0, 4.0, s0
-; GFX10-NEXT: global_store_dword v[0:1], v0, off
-; GFX10-NEXT: s_endpgm
-;
-; GFX11-LABEL: global_atomic_fadd_ret_f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_mov_b32 s4, exec_lo
-; GFX11-NEXT: s_mov_b32 s3, 0
-; GFX11-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX11-NEXT: s_mov_b32 s2, exec_lo
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: v_cmpx_eq_u32_e32 0, v0
-; GFX11-NEXT: s_cbranch_execz .LBB0_4
-; GFX11-NEXT: ; %bb.1:
-; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT: s_bcnt1_i32_b32 s4, s4
-; GFX11-NEXT: v_mov_b32_e32 v3, 0
-; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
-; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: s_load_b32 s5, s[0:1], 0x0
-; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: v_dual_mul_f32 v2, 4.0, v1 :: v_dual_mov_b32 v1, s5
-; GFX11-NEXT: .LBB0_2: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: v_mov_b32_e32 v5, v1
-; GFX11-NEXT: v_add_f32_e32 v4, v5, v2
-; GFX11-NEXT: global_atomic_cmpswap_b32 v1, v3, v[4:5], s[0:1] glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v5
-; GFX11-NEXT: s_or_b32 s3, vcc_lo, s3
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s3
-; GFX11-NEXT: s_cbranch_execnz .LBB0_2
-; GFX11-NEXT: ; %bb.3: ; %Flow
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s3
-; GFX11-NEXT: .LBB0_4: ; %Flow1
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s2
-; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX11-NEXT: v_readfirstlane_b32 s0, v1
-; GFX11-NEXT: v_mul_f32_e32 v0, 4.0, v0
-; GFX11-NEXT: v_add_f32_e32 v0, s0, v0
-; GFX11-NEXT: global_store_b32 v[0:1], v0, off
-; GFX11-NEXT: s_nop 0
-; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX11-NEXT: s_endpgm
-;
-; GFX12-LABEL: global_atomic_fadd_ret_f32:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_mov_b32 s4, exec_lo
-; GFX12-NEXT: s_mov_b32 s3, 0
-; GFX12-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX12-NEXT: s_mov_b32 s2, exec_lo
-; GFX12-NEXT: ; implicit-def: $vgpr1
-; GFX12-NEXT: v_cmpx_eq_u32_e32 0, v0
-; GFX12-NEXT: s_cbranch_execz .LBB0_4
-; GFX12-NEXT: ; %bb.1:
-; GFX12-NEXT: s_load_b64 s[0:1], s[0:1], 0x24
-; GFX12-NEXT: s_bcnt1_i32_b32 s4, s4
-; GFX12-NEXT: v_mov_b32_e32 v3, 0
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: s_load_b32 s5, s[0:1], 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: v_dual_mul_f32 v2, 4.0, v1 :: v_dual_mov_b32 v1, s5
-; GFX12-NEXT: .LBB0_2: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: v_mov_b32_e32 v5, v1
-; GFX12-NEXT: v_add_f32_e32 v4, v5, v2
-; GFX12-NEXT: global_atomic_cmpswap_b32 v1, v3, v[4:5], s[0:1] th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SYS
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v5
-; GFX12-NEXT: s_or_b32 s3, vcc_lo, s3
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s3
-; GFX12-NEXT: s_cbranch_execnz .LBB0_2
-; GFX12-NEXT: ; %bb.3: ; %Flow
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s3
-; GFX12-NEXT: .LBB0_4: ; %Flow1
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s2
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX12-NEXT: v_readfirstlane_b32 s0, v1
-; GFX12-NEXT: v_mul_f32_e32 v0, 4.0, v0
-; GFX12-NEXT: v_add_f32_e32 v0, s0, v0
-; GFX12-NEXT: global_store_b32 v[0:1], v0, off
-; GFX12-NEXT: s_nop 0
-; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX12-NEXT: s_endpgm
- %result = atomicrmw fadd ptr addrspace(1) %ptr, float 4.0 seq_cst
- store float %result, ptr addrspace(1) undef
- ret void
-}
-
-define amdgpu_kernel void @global_atomic_fadd_ret_f32_ieee(ptr addrspace(1) %ptr) #2 {
-; GFX900-LABEL: global_atomic_fadd_ret_f32_ieee:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_mov_b64 s[4:5], exec
-; GFX900-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX900-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX900-NEXT: ; implicit-def: $vgpr1
-; GFX900-NEXT: s_and_saveexec_b64 s[2:3], vcc
-; GFX900-NEXT: s_cbranch_execz .LBB1_4
-; GFX900-NEXT: ; %bb.1:
-; GFX900-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX900-NEXT: s_bcnt1_i32_b64 s7, s[4:5]
-; GFX900-NEXT: v_cvt_f32_ubyte0_e32 v1, s7
-; GFX900-NEXT: s_mov_b64 s[4:5], 0
-; GFX900-NEXT: v_mul_f32_e32 v2, 4.0, v1
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: s_load_dword s6, s[0:1], 0x0
-; GFX900-NEXT: v_mov_b32_e32 v3, 0
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v1, s6
-; GFX900-NEXT: .LBB1_2: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: v_mov_b32_e32 v5, v1
-; GFX900-NEXT: v_add_f32_e32 v4, v5, v2
-; GFX900-NEXT: global_atomic_cmpswap v1, v3, v[4:5], s[0:1] glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v1, v5
-; GFX900-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_cbranch_execnz .LBB1_2
-; GFX900-NEXT: ; %bb.3: ; %Flow
-; GFX900-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX900-NEXT: .LBB1_4: ; %Flow1
-; GFX900-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX900-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX900-NEXT: v_readfirstlane_b32 s0, v1
-; GFX900-NEXT: v_mul_f32_e32 v0, 4.0, v0
-; GFX900-NEXT: v_add_f32_e32 v0, s0, v0
-; GFX900-NEXT: global_store_dword v[0:1], v0, off
-; GFX900-NEXT: s_endpgm
-;
-; GFX908-LABEL: global_atomic_fadd_ret_f32_ieee:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_mov_b64 s[4:5], exec
-; GFX908-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX908-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX908-NEXT: ; implicit-def: $vgpr1
-; GFX908-NEXT: s_and_saveexec_b64 s[2:3], vcc
-; GFX908-NEXT: s_cbranch_execz .LBB1_4
-; GFX908-NEXT: ; %bb.1:
-; GFX908-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX908-NEXT: s_bcnt1_i32_b64 s7, s[4:5]
-; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v1, s7
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: v_mul_f32_e32 v2, 4.0, v1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: s_load_dword s6, s[0:1], 0x0
-; GFX908-NEXT: v_mov_b32_e32 v3, 0
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v1, s6
-; GFX908-NEXT: .LBB1_2: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: v_mov_b32_e32 v5, v1
-; GFX908-NEXT: v_add_f32_e32 v4, v5, v2
-; GFX908-NEXT: global_atomic_cmpswap v1, v3, v[4:5], s[0:1] glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v5
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB1_2
-; GFX908-NEXT: ; %bb.3: ; %Flow
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: .LBB1_4: ; %Flow1
-; GFX908-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX908-NEXT: v_readfirstlane_b32 s0, v1
-; GFX908-NEXT: v_mul_f32_e32 v0, 4.0, v0
-; GFX908-NEXT: v_add_f32_e32 v0, s0, v0
-; GFX908-NEXT: global_store_dword v[0:1], v0, off
-; GFX908-NEXT: s_endpgm
-;
-; GFX90A-LABEL: global_atomic_fadd_ret_f32_ieee:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_mov_b64 s[4:5], exec
-; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX90A-NEXT: ; implicit-def: $vgpr1
-; GFX90A-NEXT: s_and_saveexec_b64 s[2:3], vcc
-; GFX90A-NEXT: s_cbranch_execz .LBB1_2
-; GFX90A-NEXT: ; %bb.1:
-; GFX90A-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX90A-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
-; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v2, s4
-; GFX90A-NEXT: v_mov_b32_e32 v1, 0
-; GFX90A-NEXT: v_mul_f32_e32 v2, 4.0, v2
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: global_atomic_add_f32 v1, v1, v2, s[0:1] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: .LBB1_2:
-; GFX90A-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX90A-NEXT: v_readfirstlane_b32 s0, v1
-; GFX90A-NEXT: v_mul_f32_e32 v0, 4.0, v0
-; GFX90A-NEXT: v_add_f32_e32 v0, s0, v0
-; GFX90A-NEXT: global_store_dword v[0:1], v0, off
-; GFX90A-NEXT: s_endpgm
-;
-; GFX940-LABEL: global_atomic_fadd_ret_f32_ieee:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_mov_b64 s[4:5], exec
-; GFX940-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX940-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX940-NEXT: ; implicit-def: $vgpr1
-; GFX940-NEXT: s_and_saveexec_b64 s[2:3], vcc
-; GFX940-NEXT: s_cbranch_execz .LBB1_2
-; GFX940-NEXT: ; %bb.1:
-; GFX940-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX940-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
-; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v2, s4
-; GFX940-NEXT: v_mov_b32_e32 v1, 0
-; GFX940-NEXT: v_mul_f32_e32 v2, 4.0, v2
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: global_atomic_add_f32 v1, v1, v2, s[0:1] sc0
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: .LBB1_2:
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX940-NEXT: v_readfirstlane_b32 s0, v1
-; GFX940-NEXT: v_mul_f32_e32 v0, 4.0, v0
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_add_f32_e32 v0, s0, v0
-; GFX940-NEXT: global_store_dword v[0:1], v0, off sc0 sc1
-; GFX940-NEXT: s_endpgm
-;
-; GFX10-LABEL: global_atomic_fadd_ret_f32_ieee:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_mov_b32 s4, exec_lo
-; GFX10-NEXT: s_mov_b32 s3, 0
-; GFX10-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX10-NEXT: ; implicit-def: $vgpr1
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
-; GFX10-NEXT: s_and_saveexec_b32 s2, vcc_lo
-; GFX10-NEXT: s_cbranch_execz .LBB1_4
-; GFX10-NEXT: ; %bb.1:
-; GFX10-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX10-NEXT: s_bcnt1_i32_b32 s4, s4
-; GFX10-NEXT: v_mov_b32_e32 v3, 0
-; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
-; GFX10-NEXT: v_mul_f32_e32 v2, 4.0, v1
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: s_load_dword s5, s[0:1], 0x0
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v1, s5
-; GFX10-NEXT: .LBB1_2: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: v_mov_b32_e32 v5, v1
-; GFX10-NEXT: v_add_f32_e32 v4, v5, v2
-; GFX10-NEXT: global_atomic_cmpswap v1, v3, v[4:5], s[0:1] glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v5
-; GFX10-NEXT: s_or_b32 s3, vcc_lo, s3
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s3
-; GFX10-NEXT: s_cbranch_execnz .LBB1_2
-; GFX10-NEXT: ; %bb.3: ; %Flow
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s3
-; GFX10-NEXT: .LBB1_4: ; %Flow1
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s2
-; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX10-NEXT: v_readfirstlane_b32 s0, v1
-; GFX10-NEXT: v_mul_f32_e32 v0, 4.0, v0
-; GFX10-NEXT: v_add_f32_e32 v0, s0, v0
-; GFX10-NEXT: global_store_dword v[0:1], v0, off
-; GFX10-NEXT: s_endpgm
-;
-; GFX11-LABEL: global_atomic_fadd_ret_f32_ieee:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_mov_b32 s3, exec_lo
-; GFX11-NEXT: s_mov_b32 s2, exec_lo
-; GFX11-NEXT: v_mbcnt_lo_u32_b32 v0, s3, 0
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: v_cmpx_eq_u32_e32 0, v0
-; GFX11-NEXT: s_cbranch_execz .LBB1_2
-; GFX11-NEXT: ; %bb.1:
-; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT: s_bcnt1_i32_b32 s3, s3
-; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v1, s3
-; GFX11-NEXT: v_dual_mov_b32 v2, 0 :: v_dual_mul_f32 v1, 4.0, v1
-; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: global_atomic_add_f32 v1, v2, v1, s[0:1] glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: .LBB1_2:
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s2
-; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX11-NEXT: v_readfirstlane_b32 s0, v1
-; GFX11-NEXT: v_mul_f32_e32 v0, 4.0, v0
-; GFX11-NEXT: v_add_f32_e32 v0, s0, v0
-; GFX11-NEXT: global_store_b32 v[0:1], v0, off
-; GFX11-NEXT: s_nop 0
-; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX11-NEXT: s_endpgm
-;
-; GFX12-LABEL: global_atomic_fadd_ret_f32_ieee:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_mov_b32 s3, exec_lo
-; GFX12-NEXT: s_mov_b32 s2, exec_lo
-; GFX12-NEXT: v_mbcnt_lo_u32_b32 v0, s3, 0
-; GFX12-NEXT: ; implicit-def: $vgpr1
-; GFX12-NEXT: v_cmpx_eq_u32_e32 0, v0
-; GFX12-NEXT: s_cbranch_execz .LBB1_2
-; GFX12-NEXT: ; %bb.1:
-; GFX12-NEXT: s_load_b64 s[0:1], s[0:1], 0x24
-; GFX12-NEXT: s_bcnt1_i32_b32 s3, s3
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v1, s3
-; GFX12-NEXT: v_dual_mov_b32 v2, 0 :: v_dual_mul_f32 v1, 4.0, v1
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_atomic_add_f32 v1, v2, v1, s[0:1] th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: .LBB1_2:
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s2
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX12-NEXT: v_readfirstlane_b32 s0, v1
-; GFX12-NEXT: v_mul_f32_e32 v0, 4.0, v0
-; GFX12-NEXT: v_add_f32_e32 v0, s0, v0
-; GFX12-NEXT: global_store_b32 v[0:1], v0, off
-; GFX12-NEXT: s_nop 0
-; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX12-NEXT: s_endpgm
- %result = atomicrmw fadd ptr addrspace(1) %ptr, float 4.0 syncscope("agent") seq_cst
- store float %result, ptr addrspace(1) undef
- ret void
-}
-
-define amdgpu_kernel void @global_atomic_fadd_noret_f32(ptr addrspace(1) %ptr) #0 {
-; GFX900-LABEL: global_atomic_fadd_noret_f32:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_mov_b64 s[2:3], exec
-; GFX900-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX900-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX900-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX900-NEXT: s_cbranch_execz .LBB2_3
-; GFX900-NEXT: ; %bb.1:
-; GFX900-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX900-NEXT: s_bcnt1_i32_b64 s5, s[2:3]
-; GFX900-NEXT: v_cvt_f32_ubyte0_e32 v0, s5
-; GFX900-NEXT: s_mov_b64 s[2:3], 0
-; GFX900-NEXT: v_mul_f32_e32 v2, 4.0, v0
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: s_load_dword s4, s[0:1], 0x0
-; GFX900-NEXT: v_mov_b32_e32 v3, 0
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v1, s4
-; GFX900-NEXT: .LBB2_2: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: v_add_f32_e32 v0, v1, v2
-; GFX900-NEXT: global_atomic_cmpswap v0, v3, v[0:1], s[0:1] glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX900-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX900-NEXT: v_mov_b32_e32 v1, v0
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX900-NEXT: s_cbranch_execnz .LBB2_2
-; GFX900-NEXT: .LBB2_3:
-; GFX900-NEXT: s_endpgm
-;
-; GFX908-LABEL: global_atomic_fadd_noret_f32:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_mov_b64 s[2:3], exec
-; GFX908-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX908-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX908-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX908-NEXT: s_cbranch_execz .LBB2_2
-; GFX908-NEXT: ; %bb.1:
-; GFX908-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX908-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
-; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
-; GFX908-NEXT: v_mov_b32_e32 v0, 0
-; GFX908-NEXT: v_mul_f32_e32 v1, 4.0, v1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: .LBB2_2:
-; GFX908-NEXT: s_endpgm
-;
-; GFX90A-LABEL: global_atomic_fadd_noret_f32:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_mov_b64 s[2:3], exec
-; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX90A-NEXT: s_cbranch_execz .LBB2_2
-; GFX90A-NEXT: ; %bb.1:
-; GFX90A-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX90A-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
-; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
-; GFX90A-NEXT: v_mov_b32_e32 v0, 0
-; GFX90A-NEXT: v_mul_f32_e32 v1, 4.0, v1
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: .LBB2_2:
-; GFX90A-NEXT: s_endpgm
-;
-; GFX940-LABEL: global_atomic_fadd_noret_f32:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_mov_b64 s[2:3], exec
-; GFX940-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX940-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX940-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX940-NEXT: s_cbranch_execz .LBB2_2
-; GFX940-NEXT: ; %bb.1:
-; GFX940-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX940-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
-; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
-; GFX940-NEXT: v_mov_b32_e32 v0, 0
-; GFX940-NEXT: v_mul_f32_e32 v1, 4.0, v1
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: .LBB2_2:
-; GFX940-NEXT: s_endpgm
-;
-; GFX10-LABEL: global_atomic_fadd_noret_f32:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_mov_b32 s3, exec_lo
-; GFX10-NEXT: s_mov_b32 s2, 0
-; GFX10-NEXT: v_mbcnt_lo_u32_b32 v0, s3, 0
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
-; GFX10-NEXT: s_and_saveexec_b32 s4, vcc_lo
-; GFX10-NEXT: s_cbranch_execz .LBB2_3
-; GFX10-NEXT: ; %bb.1:
-; GFX10-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX10-NEXT: s_bcnt1_i32_b32 s3, s3
-; GFX10-NEXT: v_mov_b32_e32 v3, 0
-; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v0, s3
-; GFX10-NEXT: v_mul_f32_e32 v2, 4.0, v0
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: s_load_dword s4, s[0:1], 0x0
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v1, s4
-; GFX10-NEXT: .LBB2_2: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: v_add_f32_e32 v0, v1, v2
-; GFX10-NEXT: global_atomic_cmpswap v0, v3, v[0:1], s[0:1] glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
-; GFX10-NEXT: v_mov_b32_e32 v1, v0
-; GFX10-NEXT: s_or_b32 s2, vcc_lo, s2
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
-; GFX10-NEXT: s_cbranch_execnz .LBB2_2
-; GFX10-NEXT: .LBB2_3:
-; GFX10-NEXT: s_endpgm
-;
-; GFX11-LABEL: global_atomic_fadd_noret_f32:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_mov_b32 s2, exec_lo
-; GFX11-NEXT: s_mov_b32 s3, exec_lo
-; GFX11-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX11-NEXT: v_cmpx_eq_u32_e32 0, v0
-; GFX11-NEXT: s_cbranch_execz .LBB2_2
-; GFX11-NEXT: ; %bb.1:
-; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT: s_bcnt1_i32_b32 s2, s2
-; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v0, s2
-; GFX11-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mul_f32 v0, 4.0, v0
-; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: global_atomic_add_f32 v1, v0, s[0:1]
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: .LBB2_2:
-; GFX11-NEXT: s_endpgm
-;
-; GFX12-LABEL: global_atomic_fadd_noret_f32:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_mov_b32 s2, exec_lo
-; GFX12-NEXT: s_mov_b32 s3, exec_lo
-; GFX12-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX12-NEXT: v_cmpx_eq_u32_e32 0, v0
-; GFX12-NEXT: s_cbranch_execz .LBB2_2
-; GFX12-NEXT: ; %bb.1:
-; GFX12-NEXT: s_load_b64 s[0:1], s[0:1], 0x24
-; GFX12-NEXT: s_bcnt1_i32_b32 s2, s2
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v0, s2
-; GFX12-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mul_f32 v0, 4.0, v0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_atomic_add_f32 v1, v0, s[0:1]
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: .LBB2_2:
-; GFX12-NEXT: s_endpgm
- %result = atomicrmw fadd ptr addrspace(1) %ptr, float 4.0 syncscope("agent") seq_cst
- ret void
-}
-
-define amdgpu_kernel void @global_atomic_fadd_noret_f32_ieee(ptr addrspace(1) %ptr) #2 {
-; GFX900-LABEL: global_atomic_fadd_noret_f32_ieee:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_mov_b64 s[2:3], exec
-; GFX900-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX900-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX900-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX900-NEXT: s_cbranch_execz .LBB3_3
-; GFX900-NEXT: ; %bb.1:
-; GFX900-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX900-NEXT: s_bcnt1_i32_b64 s5, s[2:3]
-; GFX900-NEXT: v_cvt_f32_ubyte0_e32 v0, s5
-; GFX900-NEXT: s_mov_b64 s[2:3], 0
-; GFX900-NEXT: v_mul_f32_e32 v2, 4.0, v0
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: s_load_dword s4, s[0:1], 0x0
-; GFX900-NEXT: v_mov_b32_e32 v3, 0
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v1, s4
-; GFX900-NEXT: .LBB3_2: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: v_add_f32_e32 v0, v1, v2
-; GFX900-NEXT: global_atomic_cmpswap v0, v3, v[0:1], s[0:1] glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX900-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX900-NEXT: v_mov_b32_e32 v1, v0
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX900-NEXT: s_cbranch_execnz .LBB3_2
-; GFX900-NEXT: .LBB3_3:
-; GFX900-NEXT: s_endpgm
-;
-; GFX908-LABEL: global_atomic_fadd_noret_f32_ieee:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_mov_b64 s[2:3], exec
-; GFX908-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX908-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX908-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX908-NEXT: s_cbranch_execz .LBB3_2
-; GFX908-NEXT: ; %bb.1:
-; GFX908-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX908-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
-; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
-; GFX908-NEXT: v_mov_b32_e32 v0, 0
-; GFX908-NEXT: v_mul_f32_e32 v1, 4.0, v1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: .LBB3_2:
-; GFX908-NEXT: s_endpgm
-;
-; GFX90A-LABEL: global_atomic_fadd_noret_f32_ieee:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_mov_b64 s[2:3], exec
-; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX90A-NEXT: s_cbranch_execz .LBB3_2
-; GFX90A-NEXT: ; %bb.1:
-; GFX90A-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX90A-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
-; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
-; GFX90A-NEXT: v_mov_b32_e32 v0, 0
-; GFX90A-NEXT: v_mul_f32_e32 v1, 4.0, v1
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: .LBB3_2:
-; GFX90A-NEXT: s_endpgm
-;
-; GFX940-LABEL: global_atomic_fadd_noret_f32_ieee:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_mov_b64 s[2:3], exec
-; GFX940-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX940-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX940-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX940-NEXT: s_cbranch_execz .LBB3_2
-; GFX940-NEXT: ; %bb.1:
-; GFX940-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX940-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
-; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
-; GFX940-NEXT: v_mov_b32_e32 v0, 0
-; GFX940-NEXT: v_mul_f32_e32 v1, 4.0, v1
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: .LBB3_2:
-; GFX940-NEXT: s_endpgm
-;
-; GFX10-LABEL: global_atomic_fadd_noret_f32_ieee:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_mov_b32 s3, exec_lo
-; GFX10-NEXT: s_mov_b32 s2, 0
-; GFX10-NEXT: v_mbcnt_lo_u32_b32 v0, s3, 0
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
-; GFX10-NEXT: s_and_saveexec_b32 s4, vcc_lo
-; GFX10-NEXT: s_cbranch_execz .LBB3_3
-; GFX10-NEXT: ; %bb.1:
-; GFX10-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX10-NEXT: s_bcnt1_i32_b32 s3, s3
-; GFX10-NEXT: v_mov_b32_e32 v3, 0
-; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v0, s3
-; GFX10-NEXT: v_mul_f32_e32 v2, 4.0, v0
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: s_load_dword s4, s[0:1], 0x0
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v1, s4
-; GFX10-NEXT: .LBB3_2: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: v_add_f32_e32 v0, v1, v2
-; GFX10-NEXT: global_atomic_cmpswap v0, v3, v[0:1], s[0:1] glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
-; GFX10-NEXT: v_mov_b32_e32 v1, v0
-; GFX10-NEXT: s_or_b32 s2, vcc_lo, s2
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
-; GFX10-NEXT: s_cbranch_execnz .LBB3_2
-; GFX10-NEXT: .LBB3_3:
-; GFX10-NEXT: s_endpgm
-;
-; GFX11-LABEL: global_atomic_fadd_noret_f32_ieee:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_mov_b32 s2, exec_lo
-; GFX11-NEXT: s_mov_b32 s3, exec_lo
-; GFX11-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX11-NEXT: v_cmpx_eq_u32_e32 0, v0
-; GFX11-NEXT: s_cbranch_execz .LBB3_2
-; GFX11-NEXT: ; %bb.1:
-; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT: s_bcnt1_i32_b32 s2, s2
-; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v0, s2
-; GFX11-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mul_f32 v0, 4.0, v0
-; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: global_atomic_add_f32 v1, v0, s[0:1]
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: .LBB3_2:
-; GFX11-NEXT: s_endpgm
-;
-; GFX12-LABEL: global_atomic_fadd_noret_f32_ieee:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_mov_b32 s2, exec_lo
-; GFX12-NEXT: s_mov_b32 s3, exec_lo
-; GFX12-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX12-NEXT: v_cmpx_eq_u32_e32 0, v0
-; GFX12-NEXT: s_cbranch_execz .LBB3_2
-; GFX12-NEXT: ; %bb.1:
-; GFX12-NEXT: s_load_b64 s[0:1], s[0:1], 0x24
-; GFX12-NEXT: s_bcnt1_i32_b32 s2, s2
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v0, s2
-; GFX12-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mul_f32 v0, 4.0, v0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_atomic_add_f32 v1, v0, s[0:1]
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: .LBB3_2:
-; GFX12-NEXT: s_endpgm
- %result = atomicrmw fadd ptr addrspace(1) %ptr, float 4.0 syncscope("agent") seq_cst
- ret void
-}
-
-define amdgpu_kernel void @global_atomic_fadd_ret_f32_agent(ptr addrspace(1) %ptr) #0 {
-; GFX900-LABEL: global_atomic_fadd_ret_f32_agent:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_mov_b64 s[4:5], exec
-; GFX900-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX900-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX900-NEXT: ; implicit-def: $vgpr1
-; GFX900-NEXT: s_and_saveexec_b64 s[2:3], vcc
-; GFX900-NEXT: s_cbranch_execz .LBB4_4
-; GFX900-NEXT: ; %bb.1:
-; GFX900-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX900-NEXT: s_bcnt1_i32_b64 s7, s[4:5]
-; GFX900-NEXT: v_cvt_f32_ubyte0_e32 v1, s7
-; GFX900-NEXT: s_mov_b64 s[4:5], 0
-; GFX900-NEXT: v_mul_f32_e32 v2, 4.0, v1
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: s_load_dword s6, s[0:1], 0x0
-; GFX900-NEXT: v_mov_b32_e32 v3, 0
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v1, s6
-; GFX900-NEXT: .LBB4_2: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: v_mov_b32_e32 v5, v1
-; GFX900-NEXT: v_add_f32_e32 v4, v5, v2
-; GFX900-NEXT: global_atomic_cmpswap v1, v3, v[4:5], s[0:1] glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v1, v5
-; GFX900-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_cbranch_execnz .LBB4_2
-; GFX900-NEXT: ; %bb.3: ; %Flow
-; GFX900-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX900-NEXT: .LBB4_4: ; %Flow1
-; GFX900-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX900-NEXT: v_readfirstlane_b32 s0, v1
-; GFX900-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX900-NEXT: v_mad_f32 v0, v0, 4.0, s0
-; GFX900-NEXT: global_store_dword v[0:1], v0, off
-; GFX900-NEXT: s_endpgm
-;
-; GFX908-LABEL: global_atomic_fadd_ret_f32_agent:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_mov_b64 s[4:5], exec
-; GFX908-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX908-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX908-NEXT: ; implicit-def: $vgpr1
-; GFX908-NEXT: s_and_saveexec_b64 s[2:3], vcc
-; GFX908-NEXT: s_cbranch_execz .LBB4_4
-; GFX908-NEXT: ; %bb.1:
-; GFX908-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX908-NEXT: s_bcnt1_i32_b64 s7, s[4:5]
-; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v1, s7
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: v_mul_f32_e32 v2, 4.0, v1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: s_load_dword s6, s[0:1], 0x0
-; GFX908-NEXT: v_mov_b32_e32 v3, 0
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v1, s6
-; GFX908-NEXT: .LBB4_2: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: v_mov_b32_e32 v5, v1
-; GFX908-NEXT: v_add_f32_e32 v4, v5, v2
-; GFX908-NEXT: global_atomic_cmpswap v1, v3, v[4:5], s[0:1] glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v5
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB4_2
-; GFX908-NEXT: ; %bb.3: ; %Flow
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: .LBB4_4: ; %Flow1
-; GFX908-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX908-NEXT: v_readfirstlane_b32 s0, v1
-; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX908-NEXT: v_mad_f32 v0, v0, 4.0, s0
-; GFX908-NEXT: global_store_dword v[0:1], v0, off
-; GFX908-NEXT: s_endpgm
-;
-; GFX90A-LABEL: global_atomic_fadd_ret_f32_agent:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_mov_b64 s[4:5], exec
-; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX90A-NEXT: ; implicit-def: $vgpr1
-; GFX90A-NEXT: s_and_saveexec_b64 s[2:3], vcc
-; GFX90A-NEXT: s_cbranch_execz .LBB4_2
-; GFX90A-NEXT: ; %bb.1:
-; GFX90A-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX90A-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
-; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v2, s4
-; GFX90A-NEXT: v_mov_b32_e32 v1, 0
-; GFX90A-NEXT: v_mul_f32_e32 v2, 4.0, v2
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: global_atomic_add_f32 v1, v1, v2, s[0:1] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: .LBB4_2:
-; GFX90A-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX90A-NEXT: v_readfirstlane_b32 s0, v1
-; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX90A-NEXT: v_mad_f32 v0, v0, 4.0, s0
-; GFX90A-NEXT: global_store_dword v[0:1], v0, off
-; GFX90A-NEXT: s_endpgm
-;
-; GFX940-LABEL: global_atomic_fadd_ret_f32_agent:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_mov_b64 s[4:5], exec
-; GFX940-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX940-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX940-NEXT: ; implicit-def: $vgpr1
-; GFX940-NEXT: s_and_saveexec_b64 s[2:3], vcc
-; GFX940-NEXT: s_cbranch_execz .LBB4_2
-; GFX940-NEXT: ; %bb.1:
-; GFX940-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX940-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
-; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v2, s4
-; GFX940-NEXT: v_mov_b32_e32 v1, 0
-; GFX940-NEXT: v_mul_f32_e32 v2, 4.0, v2
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: global_atomic_add_f32 v1, v1, v2, s[0:1] sc0
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: .LBB4_2:
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX940-NEXT: v_readfirstlane_b32 s0, v1
-; GFX940-NEXT: v_mul_f32_e32 v0, 4.0, v0
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_add_f32_e32 v0, s0, v0
-; GFX940-NEXT: global_store_dword v[0:1], v0, off sc0 sc1
-; GFX940-NEXT: s_endpgm
-;
-; GFX10-LABEL: global_atomic_fadd_ret_f32_agent:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_mov_b32 s4, exec_lo
-; GFX10-NEXT: s_mov_b32 s3, 0
-; GFX10-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX10-NEXT: ; implicit-def: $vgpr1
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
-; GFX10-NEXT: s_and_saveexec_b32 s2, vcc_lo
-; GFX10-NEXT: s_cbranch_execz .LBB4_4
-; GFX10-NEXT: ; %bb.1:
-; GFX10-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX10-NEXT: s_bcnt1_i32_b32 s4, s4
-; GFX10-NEXT: v_mov_b32_e32 v3, 0
-; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
-; GFX10-NEXT: v_mul_f32_e32 v2, 4.0, v1
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: s_load_dword s5, s[0:1], 0x0
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v1, s5
-; GFX10-NEXT: .LBB4_2: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: v_mov_b32_e32 v5, v1
-; GFX10-NEXT: v_add_f32_e32 v4, v5, v2
-; GFX10-NEXT: global_atomic_cmpswap v1, v3, v[4:5], s[0:1] glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v5
-; GFX10-NEXT: s_or_b32 s3, vcc_lo, s3
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s3
-; GFX10-NEXT: s_cbranch_execnz .LBB4_2
-; GFX10-NEXT: ; %bb.3: ; %Flow
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s3
-; GFX10-NEXT: .LBB4_4: ; %Flow1
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s2
-; GFX10-NEXT: v_readfirstlane_b32 s0, v1
-; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX10-NEXT: v_mad_f32 v0, v0, 4.0, s0
-; GFX10-NEXT: global_store_dword v[0:1], v0, off
-; GFX10-NEXT: s_endpgm
-;
-; GFX11-LABEL: global_atomic_fadd_ret_f32_agent:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_mov_b32 s3, exec_lo
-; GFX11-NEXT: s_mov_b32 s2, exec_lo
-; GFX11-NEXT: v_mbcnt_lo_u32_b32 v0, s3, 0
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: v_cmpx_eq_u32_e32 0, v0
-; GFX11-NEXT: s_cbranch_execz .LBB4_2
-; GFX11-NEXT: ; %bb.1:
-; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT: s_bcnt1_i32_b32 s3, s3
-; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v1, s3
-; GFX11-NEXT: v_dual_mov_b32 v2, 0 :: v_dual_mul_f32 v1, 4.0, v1
-; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: global_atomic_add_f32 v1, v2, v1, s[0:1] glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: .LBB4_2:
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s2
-; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX11-NEXT: v_readfirstlane_b32 s0, v1
-; GFX11-NEXT: v_mul_f32_e32 v0, 4.0, v0
-; GFX11-NEXT: v_add_f32_e32 v0, s0, v0
-; GFX11-NEXT: global_store_b32 v[0:1], v0, off
-; GFX11-NEXT: s_nop 0
-; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX11-NEXT: s_endpgm
-;
-; GFX12-LABEL: global_atomic_fadd_ret_f32_agent:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_mov_b32 s3, exec_lo
-; GFX12-NEXT: s_mov_b32 s2, exec_lo
-; GFX12-NEXT: v_mbcnt_lo_u32_b32 v0, s3, 0
-; GFX12-NEXT: ; implicit-def: $vgpr1
-; GFX12-NEXT: v_cmpx_eq_u32_e32 0, v0
-; GFX12-NEXT: s_cbranch_execz .LBB4_2
-; GFX12-NEXT: ; %bb.1:
-; GFX12-NEXT: s_load_b64 s[0:1], s[0:1], 0x24
-; GFX12-NEXT: s_bcnt1_i32_b32 s3, s3
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v1, s3
-; GFX12-NEXT: v_dual_mov_b32 v2, 0 :: v_dual_mul_f32 v1, 4.0, v1
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_atomic_add_f32 v1, v2, v1, s[0:1] th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: .LBB4_2:
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s2
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX12-NEXT: v_readfirstlane_b32 s0, v1
-; GFX12-NEXT: v_mul_f32_e32 v0, 4.0, v0
-; GFX12-NEXT: v_add_f32_e32 v0, s0, v0
-; GFX12-NEXT: global_store_b32 v[0:1], v0, off
-; GFX12-NEXT: s_nop 0
-; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX12-NEXT: s_endpgm
- %result = atomicrmw fadd ptr addrspace(1) %ptr, float 4.0 syncscope("agent") seq_cst
- store float %result, ptr addrspace(1) undef
- ret void
-}
-
-define amdgpu_kernel void @global_atomic_fadd_ret_f32_system(ptr addrspace(1) %ptr) #0 {
-; GFX900-LABEL: global_atomic_fadd_ret_f32_system:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_mov_b64 s[4:5], exec
-; GFX900-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX900-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX900-NEXT: ; implicit-def: $vgpr1
-; GFX900-NEXT: s_and_saveexec_b64 s[2:3], vcc
-; GFX900-NEXT: s_cbranch_execz .LBB5_4
-; GFX900-NEXT: ; %bb.1:
-; GFX900-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX900-NEXT: s_bcnt1_i32_b64 s7, s[4:5]
-; GFX900-NEXT: v_cvt_f32_ubyte0_e32 v1, s7
-; GFX900-NEXT: s_mov_b64 s[4:5], 0
-; GFX900-NEXT: v_mul_f32_e32 v2, 4.0, v1
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: s_load_dword s6, s[0:1], 0x0
-; GFX900-NEXT: v_mov_b32_e32 v3, 0
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v1, s6
-; GFX900-NEXT: .LBB5_2: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: v_mov_b32_e32 v5, v1
-; GFX900-NEXT: v_add_f32_e32 v4, v5, v2
-; GFX900-NEXT: global_atomic_cmpswap v1, v3, v[4:5], s[0:1] glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v1, v5
-; GFX900-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_cbranch_execnz .LBB5_2
-; GFX900-NEXT: ; %bb.3: ; %Flow
-; GFX900-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX900-NEXT: .LBB5_4: ; %Flow1
-; GFX900-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX900-NEXT: v_readfirstlane_b32 s0, v1
-; GFX900-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX900-NEXT: v_mad_f32 v0, v0, 4.0, s0
-; GFX900-NEXT: global_store_dword v[0:1], v0, off
-; GFX900-NEXT: s_endpgm
-;
-; GFX908-LABEL: global_atomic_fadd_ret_f32_system:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_mov_b64 s[4:5], exec
-; GFX908-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX908-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX908-NEXT: ; implicit-def: $vgpr1
-; GFX908-NEXT: s_and_saveexec_b64 s[2:3], vcc
-; GFX908-NEXT: s_cbranch_execz .LBB5_4
-; GFX908-NEXT: ; %bb.1:
-; GFX908-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX908-NEXT: s_bcnt1_i32_b64 s7, s[4:5]
-; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v1, s7
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: v_mul_f32_e32 v2, 4.0, v1
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: s_load_dword s6, s[0:1], 0x0
-; GFX908-NEXT: v_mov_b32_e32 v3, 0
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v1, s6
-; GFX908-NEXT: .LBB5_2: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: v_mov_b32_e32 v5, v1
-; GFX908-NEXT: v_add_f32_e32 v4, v5, v2
-; GFX908-NEXT: global_atomic_cmpswap v1, v3, v[4:5], s[0:1] glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v5
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB5_2
-; GFX908-NEXT: ; %bb.3: ; %Flow
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: .LBB5_4: ; %Flow1
-; GFX908-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX908-NEXT: v_readfirstlane_b32 s0, v1
-; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX908-NEXT: v_mad_f32 v0, v0, 4.0, s0
-; GFX908-NEXT: global_store_dword v[0:1], v0, off
-; GFX908-NEXT: s_endpgm
-;
-; GFX90A-LABEL: global_atomic_fadd_ret_f32_system:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_mov_b64 s[4:5], exec
-; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX90A-NEXT: ; implicit-def: $vgpr1
-; GFX90A-NEXT: s_and_saveexec_b64 s[2:3], vcc
-; GFX90A-NEXT: s_cbranch_execz .LBB5_4
-; GFX90A-NEXT: ; %bb.1:
-; GFX90A-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX90A-NEXT: s_bcnt1_i32_b64 s7, s[4:5]
-; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v1, s7
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: v_mul_f32_e32 v2, 4.0, v1
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: s_load_dword s6, s[0:1], 0x0
-; GFX90A-NEXT: v_mov_b32_e32 v3, 0
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v1, s6
-; GFX90A-NEXT: .LBB5_2: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: v_mov_b32_e32 v5, v1
-; GFX90A-NEXT: v_add_f32_e32 v4, v5, v2
-; GFX90A-NEXT: buffer_wbl2
-; GFX90A-NEXT: global_atomic_cmpswap v1, v3, v[4:5], s[0:1] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_invl2
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v5
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB5_2
-; GFX90A-NEXT: ; %bb.3: ; %Flow
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: .LBB5_4: ; %Flow1
-; GFX90A-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX90A-NEXT: v_readfirstlane_b32 s0, v1
-; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX90A-NEXT: v_mad_f32 v0, v0, 4.0, s0
-; GFX90A-NEXT: global_store_dword v[0:1], v0, off
-; GFX90A-NEXT: s_endpgm
-;
-; GFX940-LABEL: global_atomic_fadd_ret_f32_system:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_mov_b64 s[4:5], exec
-; GFX940-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX940-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX940-NEXT: ; implicit-def: $vgpr1
-; GFX940-NEXT: s_and_saveexec_b64 s[2:3], vcc
-; GFX940-NEXT: s_cbranch_execz .LBB5_2
-; GFX940-NEXT: ; %bb.1:
-; GFX940-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX940-NEXT: s_bcnt1_i32_b64 s4, s[4:5]
-; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v2, s4
-; GFX940-NEXT: v_mov_b32_e32 v1, 0
-; GFX940-NEXT: v_mul_f32_e32 v2, 4.0, v2
-; GFX940-NEXT: buffer_wbl2 sc0 sc1
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: global_atomic_add_f32 v1, v1, v2, s[0:1] sc0 sc1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc0 sc1
-; GFX940-NEXT: .LBB5_2:
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX940-NEXT: v_readfirstlane_b32 s0, v1
-; GFX940-NEXT: v_mul_f32_e32 v0, 4.0, v0
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_add_f32_e32 v0, s0, v0
-; GFX940-NEXT: global_store_dword v[0:1], v0, off sc0 sc1
-; GFX940-NEXT: s_endpgm
-;
-; GFX10-LABEL: global_atomic_fadd_ret_f32_system:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_mov_b32 s4, exec_lo
-; GFX10-NEXT: s_mov_b32 s3, 0
-; GFX10-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX10-NEXT: ; implicit-def: $vgpr1
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
-; GFX10-NEXT: s_and_saveexec_b32 s2, vcc_lo
-; GFX10-NEXT: s_cbranch_execz .LBB5_4
-; GFX10-NEXT: ; %bb.1:
-; GFX10-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX10-NEXT: s_bcnt1_i32_b32 s4, s4
-; GFX10-NEXT: v_mov_b32_e32 v3, 0
-; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
-; GFX10-NEXT: v_mul_f32_e32 v2, 4.0, v1
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: s_load_dword s5, s[0:1], 0x0
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v1, s5
-; GFX10-NEXT: .LBB5_2: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: v_mov_b32_e32 v5, v1
-; GFX10-NEXT: v_add_f32_e32 v4, v5, v2
-; GFX10-NEXT: global_atomic_cmpswap v1, v3, v[4:5], s[0:1] glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v5
-; GFX10-NEXT: s_or_b32 s3, vcc_lo, s3
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s3
-; GFX10-NEXT: s_cbranch_execnz .LBB5_2
-; GFX10-NEXT: ; %bb.3: ; %Flow
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s3
-; GFX10-NEXT: .LBB5_4: ; %Flow1
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s2
-; GFX10-NEXT: v_readfirstlane_b32 s0, v1
-; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX10-NEXT: v_mad_f32 v0, v0, 4.0, s0
-; GFX10-NEXT: global_store_dword v[0:1], v0, off
-; GFX10-NEXT: s_endpgm
-;
-; GFX11-LABEL: global_atomic_fadd_ret_f32_system:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_mov_b32 s4, exec_lo
-; GFX11-NEXT: s_mov_b32 s3, 0
-; GFX11-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX11-NEXT: s_mov_b32 s2, exec_lo
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: v_cmpx_eq_u32_e32 0, v0
-; GFX11-NEXT: s_cbranch_execz .LBB5_4
-; GFX11-NEXT: ; %bb.1:
-; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT: s_bcnt1_i32_b32 s4, s4
-; GFX11-NEXT: v_mov_b32_e32 v3, 0
-; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
-; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: s_load_b32 s5, s[0:1], 0x0
-; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: v_dual_mul_f32 v2, 4.0, v1 :: v_dual_mov_b32 v1, s5
-; GFX11-NEXT: .LBB5_2: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: v_mov_b32_e32 v5, v1
-; GFX11-NEXT: v_add_f32_e32 v4, v5, v2
-; GFX11-NEXT: global_atomic_cmpswap_b32 v1, v3, v[4:5], s[0:1] glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v5
-; GFX11-NEXT: s_or_b32 s3, vcc_lo, s3
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s3
-; GFX11-NEXT: s_cbranch_execnz .LBB5_2
-; GFX11-NEXT: ; %bb.3: ; %Flow
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s3
-; GFX11-NEXT: .LBB5_4: ; %Flow1
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s2
-; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX11-NEXT: v_readfirstlane_b32 s0, v1
-; GFX11-NEXT: v_mul_f32_e32 v0, 4.0, v0
-; GFX11-NEXT: v_add_f32_e32 v0, s0, v0
-; GFX11-NEXT: global_store_b32 v[0:1], v0, off
-; GFX11-NEXT: s_nop 0
-; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX11-NEXT: s_endpgm
-;
-; GFX12-LABEL: global_atomic_fadd_ret_f32_system:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_mov_b32 s4, exec_lo
-; GFX12-NEXT: s_mov_b32 s3, 0
-; GFX12-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX12-NEXT: s_mov_b32 s2, exec_lo
-; GFX12-NEXT: ; implicit-def: $vgpr1
-; GFX12-NEXT: v_cmpx_eq_u32_e32 0, v0
-; GFX12-NEXT: s_cbranch_execz .LBB5_4
-; GFX12-NEXT: ; %bb.1:
-; GFX12-NEXT: s_load_b64 s[0:1], s[0:1], 0x24
-; GFX12-NEXT: s_bcnt1_i32_b32 s4, s4
-; GFX12-NEXT: v_mov_b32_e32 v3, 0
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v1, s4
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: s_load_b32 s5, s[0:1], 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: v_dual_mul_f32 v2, 4.0, v1 :: v_dual_mov_b32 v1, s5
-; GFX12-NEXT: .LBB5_2: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: v_mov_b32_e32 v5, v1
-; GFX12-NEXT: v_add_f32_e32 v4, v5, v2
-; GFX12-NEXT: global_atomic_cmpswap_b32 v1, v3, v[4:5], s[0:1] th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SYS
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v5
-; GFX12-NEXT: s_or_b32 s3, vcc_lo, s3
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s3
-; GFX12-NEXT: s_cbranch_execnz .LBB5_2
-; GFX12-NEXT: ; %bb.3: ; %Flow
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s3
-; GFX12-NEXT: .LBB5_4: ; %Flow1
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s2
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX12-NEXT: v_readfirstlane_b32 s0, v1
-; GFX12-NEXT: v_mul_f32_e32 v0, 4.0, v0
-; GFX12-NEXT: v_add_f32_e32 v0, s0, v0
-; GFX12-NEXT: global_store_b32 v[0:1], v0, off
-; GFX12-NEXT: s_nop 0
-; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX12-NEXT: s_endpgm
- %result = atomicrmw fadd ptr addrspace(1) %ptr, float 4.0 syncscope("one-as") seq_cst
- store float %result, ptr addrspace(1) undef
- ret void
-}
-
-define amdgpu_kernel void @global_atomic_fadd_ret_f32_wrong_subtarget(ptr addrspace(1) %ptr) #1 {
-; GCN-LABEL: global_atomic_fadd_ret_f32_wrong_subtarget:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_mov_b64 s[4:5], exec
-; GCN-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GCN-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: s_and_saveexec_b64 s[2:3], vcc
-; GCN-NEXT: s_cbranch_execz .LBB6_4
-; GCN-NEXT: ; %bb.1:
-; GCN-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GCN-NEXT: s_bcnt1_i32_b64 s7, s[4:5]
-; GCN-NEXT: v_cvt_f32_ubyte0_e32 v1, s7
-; GCN-NEXT: s_mov_b64 s[4:5], 0
-; GCN-NEXT: v_mul_f32_e32 v2, 4.0, v1
-; GCN-NEXT: s_waitcnt lgkmcnt(0)
-; GCN-NEXT: s_load_dword s6, s[0:1], 0x0
-; GCN-NEXT: v_mov_b32_e32 v3, 0
-; GCN-NEXT: s_waitcnt lgkmcnt(0)
-; GCN-NEXT: v_mov_b32_e32 v1, s6
-; GCN-NEXT: .LBB6_2: ; %atomicrmw.start
-; GCN-NEXT: ; =>This Inner Loop Header: Depth=1
-; GCN-NEXT: v_mov_b32_e32 v5, v1
-; GCN-NEXT: v_add_f32_e32 v4, v5, v2
-; GCN-NEXT: global_atomic_cmpswap v1, v3, v[4:5], s[0:1] glc
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_wbinvl1_vol
-; GCN-NEXT: v_cmp_eq_u32_e32 vcc, v1, v5
-; GCN-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GCN-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GCN-NEXT: s_cbranch_execnz .LBB6_2
-; GCN-NEXT: ; %bb.3: ; %Flow
-; GCN-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-NEXT: .LBB6_4: ; %Flow1
-; GCN-NEXT: s_or_b64 exec, exec, s[2:3]
-; GCN-NEXT: v_readfirstlane_b32 s0, v1
-; GCN-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GCN-NEXT: v_mad_f32 v0, v0, 4.0, s0
-; GCN-NEXT: global_store_dword v[0:1], v0, off
-; GCN-NEXT: s_endpgm
-;
-; GFX11-LABEL: global_atomic_fadd_ret_f32_wrong_subtarget:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_mov_b64 s[4:5], exec
-; GFX11-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX11-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX11-NEXT: ; implicit-def: $vgpr1
-; GFX11-NEXT: s_and_saveexec_b64 s[2:3], vcc
-; GFX11-NEXT: s_cbranch_execz .LBB6_4
-; GFX11-NEXT: ; %bb.1:
-; GFX11-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX11-NEXT: s_bcnt1_i32_b64 s7, s[4:5]
-; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v1, s7
-; GFX11-NEXT: s_mov_b64 s[4:5], 0
-; GFX11-NEXT: v_mul_f32_e32 v2, 4.0, v1
-; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: s_load_dword s6, s[0:1], 0x0
-; GFX11-NEXT: v_mov_b32_e32 v3, 0
-; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v1, s6
-; GFX11-NEXT: .LBB6_2: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: v_mov_b32_e32 v5, v1
-; GFX11-NEXT: v_add_f32_e32 v4, v5, v2
-; GFX11-NEXT: global_atomic_cmpswap v1, v3, v[4:5], s[0:1] glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_wbinvl1_vol
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc, v1, v5
-; GFX11-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX11-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX11-NEXT: s_cbranch_execnz .LBB6_2
-; GFX11-NEXT: ; %bb.3: ; %Flow
-; GFX11-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX11-NEXT: .LBB6_4: ; %Flow1
-; GFX11-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX11-NEXT: v_readfirstlane_b32 s0, v1
-; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX11-NEXT: v_mad_f32 v0, v0, 4.0, s0
-; GFX11-NEXT: global_store_dword v[0:1], v0, off
-; GFX11-NEXT: s_endpgm
-;
-; GFX12-LABEL: global_atomic_fadd_ret_f32_wrong_subtarget:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_mov_b64 s[4:5], exec
-; GFX12-NEXT: v_mbcnt_lo_u32_b32 v0, s4, 0
-; GFX12-NEXT: v_mbcnt_hi_u32_b32 v0, s5, v0
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX12-NEXT: ; implicit-def: $vgpr1
-; GFX12-NEXT: s_and_saveexec_b64 s[2:3], vcc
-; GFX12-NEXT: s_cbranch_execz .LBB6_4
-; GFX12-NEXT: ; %bb.1:
-; GFX12-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX12-NEXT: s_bcnt1_i32_b64 s7, s[4:5]
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v1, s7
-; GFX12-NEXT: s_mov_b64 s[4:5], 0
-; GFX12-NEXT: v_mul_f32_e32 v2, 4.0, v1
-; GFX12-NEXT: s_waitcnt lgkmcnt(0)
-; GFX12-NEXT: s_load_dword s6, s[0:1], 0x0
-; GFX12-NEXT: v_mov_b32_e32 v3, 0
-; GFX12-NEXT: s_waitcnt lgkmcnt(0)
-; GFX12-NEXT: v_mov_b32_e32 v1, s6
-; GFX12-NEXT: .LBB6_2: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: v_mov_b32_e32 v5, v1
-; GFX12-NEXT: v_add_f32_e32 v4, v5, v2
-; GFX12-NEXT: global_atomic_cmpswap v1, v3, v[4:5], s[0:1] glc
-; GFX12-NEXT: s_waitcnt vmcnt(0)
-; GFX12-NEXT: buffer_wbinvl1_vol
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc, v1, v5
-; GFX12-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX12-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX12-NEXT: s_cbranch_execnz .LBB6_2
-; GFX12-NEXT: ; %bb.3: ; %Flow
-; GFX12-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX12-NEXT: .LBB6_4: ; %Flow1
-; GFX12-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX12-NEXT: v_readfirstlane_b32 s0, v1
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v0, v0
-; GFX12-NEXT: v_mad_f32 v0, v0, 4.0, s0
-; GFX12-NEXT: global_store_dword v[0:1], v0, off
-; GFX12-NEXT: s_endpgm
- %result = atomicrmw fadd ptr addrspace(1) %ptr, float 4.0 syncscope("agent") seq_cst
- store float %result, ptr addrspace(1) undef
- ret void
-}
-
-define amdgpu_kernel void @global_atomic_fadd_noret_f32_wrong_subtarget(ptr addrspace(1) %ptr) #1 {
-; GCN-LABEL: global_atomic_fadd_noret_f32_wrong_subtarget:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_mov_b64 s[2:3], exec
-; GCN-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GCN-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
-; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GCN-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GCN-NEXT: s_cbranch_execz .LBB7_2
-; GCN-NEXT: ; %bb.1:
-; GCN-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GCN-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
-; GCN-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
-; GCN-NEXT: v_mov_b32_e32 v0, 0
-; GCN-NEXT: v_mul_f32_e32 v1, 4.0, v1
-; GCN-NEXT: s_waitcnt lgkmcnt(0)
-; GCN-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_wbinvl1_vol
-; GCN-NEXT: .LBB7_2:
-; GCN-NEXT: s_endpgm
-;
-; GFX11-LABEL: global_atomic_fadd_noret_f32_wrong_subtarget:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_mov_b64 s[2:3], exec
-; GFX11-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX11-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX11-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX11-NEXT: s_cbranch_execz .LBB7_2
-; GFX11-NEXT: ; %bb.1:
-; GFX11-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX11-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
-; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
-; GFX11-NEXT: v_mov_b32_e32 v0, 0
-; GFX11-NEXT: v_mul_f32_e32 v1, 4.0, v1
-; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_wbinvl1_vol
-; GFX11-NEXT: .LBB7_2:
-; GFX11-NEXT: s_endpgm
-;
-; GFX12-LABEL: global_atomic_fadd_noret_f32_wrong_subtarget:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_mov_b64 s[2:3], exec
-; GFX12-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX12-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX12-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX12-NEXT: s_cbranch_execz .LBB7_2
-; GFX12-NEXT: ; %bb.1:
-; GFX12-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX12-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
-; GFX12-NEXT: v_mov_b32_e32 v0, 0
-; GFX12-NEXT: v_mul_f32_e32 v1, 4.0, v1
-; GFX12-NEXT: s_waitcnt lgkmcnt(0)
-; GFX12-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
-; GFX12-NEXT: s_waitcnt vmcnt(0)
-; GFX12-NEXT: buffer_wbinvl1_vol
-; GFX12-NEXT: .LBB7_2:
-; GFX12-NEXT: s_endpgm
- %result = atomicrmw fadd ptr addrspace(1) %ptr, float 4.0 syncscope("agent") seq_cst
- ret void
-}
-
-define amdgpu_kernel void @global_atomic_fadd_noret_f32_safe(ptr addrspace(1) %ptr) {
-; GFX900-LABEL: global_atomic_fadd_noret_f32_safe:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_mov_b64 s[2:3], exec
-; GFX900-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX900-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX900-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX900-NEXT: s_cbranch_execz .LBB8_3
-; GFX900-NEXT: ; %bb.1:
-; GFX900-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX900-NEXT: s_bcnt1_i32_b64 s5, s[2:3]
-; GFX900-NEXT: v_cvt_f32_ubyte0_e32 v0, s5
-; GFX900-NEXT: s_mov_b64 s[2:3], 0
-; GFX900-NEXT: v_mul_f32_e32 v2, 4.0, v0
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: s_load_dword s4, s[0:1], 0x0
-; GFX900-NEXT: v_mov_b32_e32 v3, 0
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v1, s4
-; GFX900-NEXT: .LBB8_2: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: v_add_f32_e32 v0, v1, v2
-; GFX900-NEXT: global_atomic_cmpswap v0, v3, v[0:1], s[0:1] glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX900-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX900-NEXT: v_mov_b32_e32 v1, v0
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX900-NEXT: s_cbranch_execnz .LBB8_2
-; GFX900-NEXT: .LBB8_3:
-; GFX900-NEXT: s_endpgm
-;
-; GFX908-LABEL: global_atomic_fadd_noret_f32_safe:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_mov_b64 s[2:3], exec
-; GFX908-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX908-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX908-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX908-NEXT: s_cbranch_execz .LBB8_3
-; GFX908-NEXT: ; %bb.1:
-; GFX908-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX908-NEXT: s_bcnt1_i32_b64 s5, s[2:3]
-; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v0, s5
-; GFX908-NEXT: s_mov_b64 s[2:3], 0
-; GFX908-NEXT: v_mul_f32_e32 v2, 4.0, v0
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: s_load_dword s4, s[0:1], 0x0
-; GFX908-NEXT: v_mov_b32_e32 v3, 0
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v1, s4
-; GFX908-NEXT: .LBB8_2: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: v_add_f32_e32 v0, v1, v2
-; GFX908-NEXT: global_atomic_cmpswap v0, v3, v[0:1], s[0:1] glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX908-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX908-NEXT: v_mov_b32_e32 v1, v0
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX908-NEXT: s_cbranch_execnz .LBB8_2
-; GFX908-NEXT: .LBB8_3:
-; GFX908-NEXT: s_endpgm
-;
-; GFX90A-LABEL: global_atomic_fadd_noret_f32_safe:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_mov_b64 s[2:3], exec
-; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX90A-NEXT: s_cbranch_execz .LBB8_3
-; GFX90A-NEXT: ; %bb.1:
-; GFX90A-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX90A-NEXT: s_bcnt1_i32_b64 s5, s[2:3]
-; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v0, s5
-; GFX90A-NEXT: s_mov_b64 s[2:3], 0
-; GFX90A-NEXT: v_mul_f32_e32 v2, 4.0, v0
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: s_load_dword s4, s[0:1], 0x0
-; GFX90A-NEXT: v_mov_b32_e32 v3, 0
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v1, s4
-; GFX90A-NEXT: .LBB8_2: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: v_add_f32_e32 v0, v1, v2
-; GFX90A-NEXT: global_atomic_cmpswap v0, v3, v[0:1], s[0:1] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX90A-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX90A-NEXT: v_mov_b32_e32 v1, v0
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX90A-NEXT: s_cbranch_execnz .LBB8_2
-; GFX90A-NEXT: .LBB8_3:
-; GFX90A-NEXT: s_endpgm
-;
-; GFX940-LABEL: global_atomic_fadd_noret_f32_safe:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_mov_b64 s[2:3], exec
-; GFX940-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX940-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX940-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX940-NEXT: s_cbranch_execz .LBB8_2
-; GFX940-NEXT: ; %bb.1:
-; GFX940-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX940-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
-; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
-; GFX940-NEXT: v_mov_b32_e32 v0, 0
-; GFX940-NEXT: v_mul_f32_e32 v1, 4.0, v1
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: .LBB8_2:
-; GFX940-NEXT: s_endpgm
-;
-; GFX10-LABEL: global_atomic_fadd_noret_f32_safe:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_mov_b32 s3, exec_lo
-; GFX10-NEXT: s_mov_b32 s2, 0
-; GFX10-NEXT: v_mbcnt_lo_u32_b32 v0, s3, 0
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
-; GFX10-NEXT: s_and_saveexec_b32 s4, vcc_lo
-; GFX10-NEXT: s_cbranch_execz .LBB8_3
-; GFX10-NEXT: ; %bb.1:
-; GFX10-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX10-NEXT: s_bcnt1_i32_b32 s3, s3
-; GFX10-NEXT: v_mov_b32_e32 v3, 0
-; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v0, s3
-; GFX10-NEXT: v_mul_f32_e32 v2, 4.0, v0
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: s_load_dword s4, s[0:1], 0x0
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v1, s4
-; GFX10-NEXT: .LBB8_2: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: v_add_f32_e32 v0, v1, v2
-; GFX10-NEXT: global_atomic_cmpswap v0, v3, v[0:1], s[0:1] glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
-; GFX10-NEXT: v_mov_b32_e32 v1, v0
-; GFX10-NEXT: s_or_b32 s2, vcc_lo, s2
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
-; GFX10-NEXT: s_cbranch_execnz .LBB8_2
-; GFX10-NEXT: .LBB8_3:
-; GFX10-NEXT: s_endpgm
-;
-; GFX11-LABEL: global_atomic_fadd_noret_f32_safe:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_mov_b32 s3, exec_lo
-; GFX11-NEXT: s_mov_b32 s2, 0
-; GFX11-NEXT: v_mbcnt_lo_u32_b32 v0, s3, 0
-; GFX11-NEXT: s_mov_b32 s4, exec_lo
-; GFX11-NEXT: v_cmpx_eq_u32_e32 0, v0
-; GFX11-NEXT: s_cbranch_execz .LBB8_3
-; GFX11-NEXT: ; %bb.1:
-; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT: s_bcnt1_i32_b32 s3, s3
-; GFX11-NEXT: v_mov_b32_e32 v3, 0
-; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v0, s3
-; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: s_load_b32 s4, s[0:1], 0x0
-; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: v_dual_mul_f32 v2, 4.0, v0 :: v_dual_mov_b32 v1, s4
-; GFX11-NEXT: .LBB8_2: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: v_add_f32_e32 v0, v1, v2
-; GFX11-NEXT: global_atomic_cmpswap_b32 v0, v3, v[0:1], s[0:1] glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
-; GFX11-NEXT: v_mov_b32_e32 v1, v0
-; GFX11-NEXT: s_or_b32 s2, vcc_lo, s2
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
-; GFX11-NEXT: s_cbranch_execnz .LBB8_2
-; GFX11-NEXT: .LBB8_3:
-; GFX11-NEXT: s_endpgm
-;
-; GFX12-LABEL: global_atomic_fadd_noret_f32_safe:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_mov_b32 s3, exec_lo
-; GFX12-NEXT: s_mov_b32 s2, 0
-; GFX12-NEXT: v_mbcnt_lo_u32_b32 v0, s3, 0
-; GFX12-NEXT: s_mov_b32 s4, exec_lo
-; GFX12-NEXT: v_cmpx_eq_u32_e32 0, v0
-; GFX12-NEXT: s_cbranch_execz .LBB8_3
-; GFX12-NEXT: ; %bb.1:
-; GFX12-NEXT: s_load_b64 s[0:1], s[0:1], 0x24
-; GFX12-NEXT: s_bcnt1_i32_b32 s3, s3
-; GFX12-NEXT: v_mov_b32_e32 v3, 0
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v0, s3
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: s_load_b32 s4, s[0:1], 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: v_dual_mul_f32 v2, 4.0, v0 :: v_dual_mov_b32 v1, s4
-; GFX12-NEXT: .LBB8_2: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: v_add_f32_e32 v0, v1, v2
-; GFX12-NEXT: global_atomic_cmpswap_b32 v0, v3, v[0:1], s[0:1] th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
-; GFX12-NEXT: v_mov_b32_e32 v1, v0
-; GFX12-NEXT: s_or_b32 s2, vcc_lo, s2
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
-; GFX12-NEXT: s_cbranch_execnz .LBB8_2
-; GFX12-NEXT: .LBB8_3:
-; GFX12-NEXT: s_endpgm
- %result = atomicrmw fadd ptr addrspace(1) %ptr, float 4.0 syncscope("agent") seq_cst
- ret void
-}
-
-define amdgpu_kernel void @infer_as_before_atomic(ptr addrspace(4) %arg) #0 {
-; GFX900-LABEL: infer_as_before_atomic:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_mov_b64 s[2:3], exec
-; GFX900-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX900-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX900-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX900-NEXT: s_cbranch_execz .LBB9_3
-; GFX900-NEXT: ; %bb.1:
-; GFX900-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX900-NEXT: s_bcnt1_i32_b64 s5, s[2:3]
-; GFX900-NEXT: s_mov_b64 s[2:3], 0
-; GFX900-NEXT: v_cvt_f32_ubyte0_e32 v2, s5
-; GFX900-NEXT: v_mov_b32_e32 v3, 0
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: s_load_dword s4, s[0:1], 0x0
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v1, s4
-; GFX900-NEXT: .LBB9_2: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: v_add_f32_e32 v0, v1, v2
-; GFX900-NEXT: global_atomic_cmpswap v0, v3, v[0:1], s[0:1] glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX900-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX900-NEXT: v_mov_b32_e32 v1, v0
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX900-NEXT: s_cbranch_execnz .LBB9_2
-; GFX900-NEXT: .LBB9_3:
-; GFX900-NEXT: s_endpgm
-;
-; GFX908-LABEL: infer_as_before_atomic:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_mov_b64 s[2:3], exec
-; GFX908-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX908-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX908-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX908-NEXT: s_cbranch_execz .LBB9_2
-; GFX908-NEXT: ; %bb.1:
-; GFX908-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX908-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
-; GFX908-NEXT: v_mov_b32_e32 v0, 0
-; GFX908-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
-; GFX908-NEXT: .LBB9_2:
-; GFX908-NEXT: s_endpgm
-;
-; GFX90A-LABEL: infer_as_before_atomic:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_mov_b64 s[2:3], exec
-; GFX90A-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX90A-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX90A-NEXT: s_cbranch_execz .LBB9_2
-; GFX90A-NEXT: ; %bb.1:
-; GFX90A-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX90A-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
-; GFX90A-NEXT: v_mov_b32_e32 v0, 0
-; GFX90A-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
-; GFX90A-NEXT: .LBB9_2:
-; GFX90A-NEXT: s_endpgm
-;
-; GFX940-LABEL: infer_as_before_atomic:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_mov_b64 s[2:3], exec
-; GFX940-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX940-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX940-NEXT: s_and_saveexec_b64 s[4:5], vcc
-; GFX940-NEXT: s_cbranch_execz .LBB9_2
-; GFX940-NEXT: ; %bb.1:
-; GFX940-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX940-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v0, 0
-; GFX940-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
-; GFX940-NEXT: .LBB9_2:
-; GFX940-NEXT: s_endpgm
-;
-; GFX10-LABEL: infer_as_before_atomic:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_mov_b32 s3, exec_lo
-; GFX10-NEXT: s_mov_b32 s2, 0
-; GFX10-NEXT: v_mbcnt_lo_u32_b32 v0, s3, 0
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
-; GFX10-NEXT: s_and_saveexec_b32 s4, vcc_lo
-; GFX10-NEXT: s_cbranch_execz .LBB9_3
-; GFX10-NEXT: ; %bb.1:
-; GFX10-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX10-NEXT: s_bcnt1_i32_b32 s3, s3
-; GFX10-NEXT: v_mov_b32_e32 v3, 0
-; GFX10-NEXT: v_cvt_f32_ubyte0_e32 v2, s3
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: s_load_dword s4, s[0:1], 0x0
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v1, s4
-; GFX10-NEXT: .LBB9_2: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: v_add_f32_e32 v0, v1, v2
-; GFX10-NEXT: global_atomic_cmpswap v0, v3, v[0:1], s[0:1] glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
-; GFX10-NEXT: v_mov_b32_e32 v1, v0
-; GFX10-NEXT: s_or_b32 s2, vcc_lo, s2
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
-; GFX10-NEXT: s_cbranch_execnz .LBB9_2
-; GFX10-NEXT: .LBB9_3:
-; GFX10-NEXT: s_endpgm
-;
-; GFX11-LABEL: infer_as_before_atomic:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_mov_b32 s2, exec_lo
-; GFX11-NEXT: s_mov_b32 s3, exec_lo
-; GFX11-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX11-NEXT: v_cmpx_eq_u32_e32 0, v0
-; GFX11-NEXT: s_cbranch_execz .LBB9_2
-; GFX11-NEXT: ; %bb.1:
-; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT: s_bcnt1_i32_b32 s2, s2
-; GFX11-NEXT: v_mov_b32_e32 v0, 0
-; GFX11-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
-; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
-; GFX11-NEXT: .LBB9_2:
-; GFX11-NEXT: s_nop 0
-; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX11-NEXT: s_endpgm
-;
-; GFX12-LABEL: infer_as_before_atomic:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_mov_b32 s2, exec_lo
-; GFX12-NEXT: s_mov_b32 s3, exec_lo
-; GFX12-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
-; GFX12-NEXT: v_cmpx_eq_u32_e32 0, v0
-; GFX12-NEXT: s_cbranch_execz .LBB9_2
-; GFX12-NEXT: ; %bb.1:
-; GFX12-NEXT: s_load_b64 s[0:1], s[0:1], 0x24
-; GFX12-NEXT: s_bcnt1_i32_b32 s2, s2
-; GFX12-NEXT: v_mov_b32_e32 v0, 0
-; GFX12-NEXT: v_cvt_f32_ubyte0_e32 v1, s2
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_atomic_add_f32 v0, v1, s[0:1]
-; GFX12-NEXT: .LBB9_2:
-; GFX12-NEXT: s_nop 0
-; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX12-NEXT: s_endpgm
- %load = load ptr, ptr addrspace(4) %arg
- %v = atomicrmw fadd ptr %load, float 1.0 syncscope("agent-one-as") monotonic, align 4
- ret void
-}
-
-define amdgpu_kernel void @global_atomic_fadd_ret_bf16_agent(ptr addrspace(1) %ptr) #0 {
-; GFX900-LABEL: global_atomic_fadd_ret_bf16_agent:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x24
-; GFX900-NEXT: s_mov_b64 s[0:1], 0
-; GFX900-NEXT: s_movk_i32 s4, 0x7fff
-; GFX900-NEXT: v_mov_b32_e32 v0, 0
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: s_and_b32 s2, s6, -4
-; GFX900-NEXT: s_mov_b32 s3, s7
-; GFX900-NEXT: s_load_dword s7, s[2:3], 0x0
-; GFX900-NEXT: s_and_b32 s5, s6, 3
-; GFX900-NEXT: s_lshl_b32 s5, s5, 3
-; GFX900-NEXT: s_lshl_b32 s6, 0xffff, s5
-; GFX900-NEXT: s_not_b32 s6, s6
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v1, s7
-; GFX900-NEXT: .LBB10_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: v_mov_b32_e32 v2, v1
-; GFX900-NEXT: v_lshrrev_b32_sdwa v1, s5, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX900-NEXT: v_add_f32_e32 v1, 4.0, v1
-; GFX900-NEXT: v_bfe_u32 v3, v1, 16, 1
-; GFX900-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX900-NEXT: v_add3_u32 v3, v3, v1, s4
-; GFX900-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX900-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
-; GFX900-NEXT: v_lshlrev_b32_sdwa v1, s5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; GFX900-NEXT: v_and_or_b32 v1, v2, s6, v1
-; GFX900-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[2:3] glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
-; GFX900-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX900-NEXT: s_cbranch_execnz .LBB10_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX900-NEXT: v_lshrrev_b32_e32 v0, s5, v1
-; GFX900-NEXT: global_store_short v[0:1], v0, off
-; GFX900-NEXT: s_endpgm
-;
-; GFX908-LABEL: global_atomic_fadd_ret_bf16_agent:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x24
-; GFX908-NEXT: s_mov_b64 s[0:1], 0
-; GFX908-NEXT: s_movk_i32 s4, 0x7fff
-; GFX908-NEXT: v_mov_b32_e32 v0, 0
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: s_and_b32 s2, s6, -4
-; GFX908-NEXT: s_mov_b32 s3, s7
-; GFX908-NEXT: s_load_dword s7, s[2:3], 0x0
-; GFX908-NEXT: s_and_b32 s5, s6, 3
-; GFX908-NEXT: s_lshl_b32 s5, s5, 3
-; GFX908-NEXT: s_lshl_b32 s6, 0xffff, s5
-; GFX908-NEXT: s_not_b32 s6, s6
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v1, s7
-; GFX908-NEXT: .LBB10_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: v_mov_b32_e32 v2, v1
-; GFX908-NEXT: v_lshrrev_b32_sdwa v1, s5, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX908-NEXT: v_add_f32_e32 v1, 4.0, v1
-; GFX908-NEXT: v_bfe_u32 v3, v1, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX908-NEXT: v_add3_u32 v3, v3, v1, s4
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX908-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
-; GFX908-NEXT: v_lshlrev_b32_sdwa v1, s5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; GFX908-NEXT: v_and_or_b32 v1, v2, s6, v1
-; GFX908-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[2:3] glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
-; GFX908-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX908-NEXT: s_cbranch_execnz .LBB10_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX908-NEXT: v_lshrrev_b32_e32 v0, s5, v1
-; GFX908-NEXT: global_store_short v[0:1], v0, off
-; GFX908-NEXT: s_endpgm
-;
-; GFX90A-LABEL: global_atomic_fadd_ret_bf16_agent:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x24
-; GFX90A-NEXT: s_mov_b64 s[0:1], 0
-; GFX90A-NEXT: s_movk_i32 s4, 0x7fff
-; GFX90A-NEXT: v_mov_b32_e32 v0, 0
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: s_and_b32 s2, s6, -4
-; GFX90A-NEXT: s_mov_b32 s3, s7
-; GFX90A-NEXT: s_load_dword s7, s[2:3], 0x0
-; GFX90A-NEXT: s_and_b32 s5, s6, 3
-; GFX90A-NEXT: s_lshl_b32 s5, s5, 3
-; GFX90A-NEXT: s_lshl_b32 s6, 0xffff, s5
-; GFX90A-NEXT: s_not_b32 s6, s6
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v1, s7
-; GFX90A-NEXT: .LBB10_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: v_mov_b32_e32 v3, v1
-; GFX90A-NEXT: v_lshrrev_b32_sdwa v1, s5, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX90A-NEXT: v_add_f32_e32 v1, 4.0, v1
-; GFX90A-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX90A-NEXT: v_add3_u32 v2, v2, v1, s4
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX90A-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc
-; GFX90A-NEXT: v_lshlrev_b32_sdwa v1, s5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; GFX90A-NEXT: v_and_or_b32 v2, v3, s6, v1
-; GFX90A-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[2:3] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v3
-; GFX90A-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX90A-NEXT: s_cbranch_execnz .LBB10_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX90A-NEXT: v_lshrrev_b32_e32 v0, s5, v1
-; GFX90A-NEXT: global_store_short v[0:1], v0, off
-; GFX90A-NEXT: s_endpgm
-;
-; GFX940-LABEL: global_atomic_fadd_ret_bf16_agent:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x24
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_mov_b32_e32 v0, 0
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: s_and_b32 s2, s6, -4
-; GFX940-NEXT: s_mov_b32 s3, s7
-; GFX940-NEXT: s_load_dword s7, s[2:3], 0x0
-; GFX940-NEXT: s_and_b32 s5, s6, 3
-; GFX940-NEXT: s_lshl_b32 s5, s5, 3
-; GFX940-NEXT: s_lshl_b32 s6, 0xffff, s5
-; GFX940-NEXT: s_not_b32 s6, s6
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v1, s7
-; GFX940-NEXT: .LBB10_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: v_mov_b32_e32 v3, v1
-; GFX940-NEXT: v_lshrrev_b32_sdwa v1, s5, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_add_f32_e32 v1, 4.0, v1
-; GFX940-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX940-NEXT: v_add3_u32 v2, v2, v1, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX940-NEXT: s_nop 1
-; GFX940-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc
-; GFX940-NEXT: v_lshlrev_b32_sdwa v1, s5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; GFX940-NEXT: v_and_or_b32 v2, v3, s6, v1
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[2:3] sc0
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v1, v3
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB10_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: v_lshrrev_b32_e32 v0, s5, v1
-; GFX940-NEXT: global_store_short v[0:1], v0, off sc0 sc1
-; GFX940-NEXT: s_endpgm
-;
-; GFX10-LABEL: global_atomic_fadd_ret_bf16_agent:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x24
-; GFX10-NEXT: v_mov_b32_e32 v0, 0
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: s_and_b32 s0, s2, -4
-; GFX10-NEXT: s_mov_b32 s1, s3
-; GFX10-NEXT: s_and_b32 s2, s2, 3
-; GFX10-NEXT: s_load_dword s3, s[0:1], 0x0
-; GFX10-NEXT: s_lshl_b32 s2, s2, 3
-; GFX10-NEXT: s_lshl_b32 s4, 0xffff, s2
-; GFX10-NEXT: s_not_b32 s4, s4
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v1, s3
-; GFX10-NEXT: s_mov_b32 s3, 0
-; GFX10-NEXT: .LBB10_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: v_mov_b32_e32 v2, v1
-; GFX10-NEXT: v_lshrrev_b32_sdwa v1, s2, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX10-NEXT: v_add_f32_e32 v1, 4.0, v1
-; GFX10-NEXT: v_bfe_u32 v3, v1, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX10-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
-; GFX10-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo
-; GFX10-NEXT: v_lshlrev_b32_sdwa v1, s2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; GFX10-NEXT: v_and_or_b32 v1, v2, s4, v1
-; GFX10-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[0:1] glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
-; GFX10-NEXT: s_or_b32 s3, vcc_lo, s3
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s3
-; GFX10-NEXT: s_cbranch_execnz .LBB10_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s3
-; GFX10-NEXT: v_lshrrev_b32_e32 v0, s2, v1
-; GFX10-NEXT: global_store_short v[0:1], v0, off
-; GFX10-NEXT: s_endpgm
-;
-; GFX11-LABEL: global_atomic_fadd_ret_bf16_agent:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_load_b64 s[2:3], s[0:1], 0x24
-; GFX11-NEXT: v_mov_b32_e32 v0, 0
-; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: s_and_b32 s0, s2, -4
-; GFX11-NEXT: s_mov_b32 s1, s3
-; GFX11-NEXT: s_and_b32 s2, s2, 3
-; GFX11-NEXT: s_load_b32 s3, s[0:1], 0x0
-; GFX11-NEXT: s_lshl_b32 s2, s2, 3
-; GFX11-NEXT: s_lshl_b32 s4, 0xffff, s2
-; GFX11-NEXT: s_not_b32 s4, s4
-; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v1, s3
-; GFX11-NEXT: s_mov_b32 s3, 0
-; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB10_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: v_mov_b32_e32 v2, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, s2, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_add_f32_e32 v1, 4.0, v1
-; GFX11-NEXT: v_bfe_u32 v3, v1, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, s2, v1
-; GFX11-NEXT: v_and_or_b32 v1, v2, s4, v1
-; GFX11-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
-; GFX11-NEXT: s_or_b32 s3, vcc_lo, s3
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s3
-; GFX11-NEXT: s_cbranch_execnz .LBB10_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s3
-; GFX11-NEXT: v_lshrrev_b32_e32 v0, s2, v1
-; GFX11-NEXT: global_store_b16 v[0:1], v0, off
-; GFX11-NEXT: s_nop 0
-; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX11-NEXT: s_endpgm
-;
-; GFX12-LABEL: global_atomic_fadd_ret_bf16_agent:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_load_b64 s[2:3], s[0:1], 0x24
-; GFX12-NEXT: v_mov_b32_e32 v0, 0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: s_and_b32 s0, s2, -4
-; GFX12-NEXT: s_mov_b32 s1, s3
-; GFX12-NEXT: s_and_b32 s2, s2, 3
-; GFX12-NEXT: s_load_b32 s3, s[0:1], 0x0
-; GFX12-NEXT: s_lshl_b32 s2, s2, 3
-; GFX12-NEXT: s_lshl_b32 s4, 0xffff, s2
-; GFX12-NEXT: s_not_b32 s4, s4
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v1, s3
-; GFX12-NEXT: s_mov_b32 s3, 0
-; GFX12-NEXT: .LBB10_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: v_mov_b32_e32 v2, v1
-; GFX12-NEXT: v_lshrrev_b32_e32 v1, s2, v2
-; GFX12-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX12-NEXT: v_add_f32_e32 v1, 4.0, v1
-; GFX12-NEXT: v_bfe_u32 v3, v1, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX12-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
-; GFX12-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo
-; GFX12-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GFX12-NEXT: v_lshlrev_b32_e32 v1, s2, v1
-; GFX12-NEXT: v_and_or_b32 v1, v2, s4, v1
-; GFX12-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
-; GFX12-NEXT: s_or_b32 s3, vcc_lo, s3
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s3
-; GFX12-NEXT: s_cbranch_execnz .LBB10_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s3
-; GFX12-NEXT: v_lshrrev_b32_e32 v0, s2, v1
-; GFX12-NEXT: global_store_b16 v[0:1], v0, off
-; GFX12-NEXT: s_nop 0
-; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX12-NEXT: s_endpgm
- %result = atomicrmw fadd ptr addrspace(1) %ptr, bfloat 4.0 syncscope("agent") seq_cst
- store bfloat %result, ptr addrspace(1) undef
- ret void
-}
-
-define amdgpu_kernel void @global_atomic_fadd_ret_bf16_system(ptr addrspace(1) %ptr) #0 {
-; GFX900-LABEL: global_atomic_fadd_ret_bf16_system:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x24
-; GFX900-NEXT: s_mov_b64 s[0:1], 0
-; GFX900-NEXT: s_movk_i32 s4, 0x7fff
-; GFX900-NEXT: v_mov_b32_e32 v0, 0
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: s_and_b32 s2, s6, -4
-; GFX900-NEXT: s_mov_b32 s3, s7
-; GFX900-NEXT: s_load_dword s7, s[2:3], 0x0
-; GFX900-NEXT: s_and_b32 s5, s6, 3
-; GFX900-NEXT: s_lshl_b32 s5, s5, 3
-; GFX900-NEXT: s_lshl_b32 s6, 0xffff, s5
-; GFX900-NEXT: s_not_b32 s6, s6
-; GFX900-NEXT: s_waitcnt lgkmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v1, s7
-; GFX900-NEXT: .LBB11_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: v_mov_b32_e32 v2, v1
-; GFX900-NEXT: v_lshrrev_b32_sdwa v1, s5, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX900-NEXT: v_add_f32_e32 v1, 4.0, v1
-; GFX900-NEXT: v_bfe_u32 v3, v1, 16, 1
-; GFX900-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX900-NEXT: v_add3_u32 v3, v3, v1, s4
-; GFX900-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX900-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
-; GFX900-NEXT: v_lshlrev_b32_sdwa v1, s5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; GFX900-NEXT: v_and_or_b32 v1, v2, s6, v1
-; GFX900-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[2:3] glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
-; GFX900-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX900-NEXT: s_cbranch_execnz .LBB11_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX900-NEXT: v_lshrrev_b32_e32 v0, s5, v1
-; GFX900-NEXT: global_store_short v[0:1], v0, off
-; GFX900-NEXT: s_endpgm
-;
-; GFX908-LABEL: global_atomic_fadd_ret_bf16_system:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x24
-; GFX908-NEXT: s_mov_b64 s[0:1], 0
-; GFX908-NEXT: s_movk_i32 s4, 0x7fff
-; GFX908-NEXT: v_mov_b32_e32 v0, 0
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: s_and_b32 s2, s6, -4
-; GFX908-NEXT: s_mov_b32 s3, s7
-; GFX908-NEXT: s_load_dword s7, s[2:3], 0x0
-; GFX908-NEXT: s_and_b32 s5, s6, 3
-; GFX908-NEXT: s_lshl_b32 s5, s5, 3
-; GFX908-NEXT: s_lshl_b32 s6, 0xffff, s5
-; GFX908-NEXT: s_not_b32 s6, s6
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v1, s7
-; GFX908-NEXT: .LBB11_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: v_mov_b32_e32 v2, v1
-; GFX908-NEXT: v_lshrrev_b32_sdwa v1, s5, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX908-NEXT: v_add_f32_e32 v1, 4.0, v1
-; GFX908-NEXT: v_bfe_u32 v3, v1, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX908-NEXT: v_add3_u32 v3, v3, v1, s4
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX908-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
-; GFX908-NEXT: v_lshlrev_b32_sdwa v1, s5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; GFX908-NEXT: v_and_or_b32 v1, v2, s6, v1
-; GFX908-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[2:3] glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2
-; GFX908-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX908-NEXT: s_cbranch_execnz .LBB11_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX908-NEXT: v_lshrrev_b32_e32 v0, s5, v1
-; GFX908-NEXT: global_store_short v[0:1], v0, off
-; GFX908-NEXT: s_endpgm
-;
-; GFX90A-LABEL: global_atomic_fadd_ret_bf16_system:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x24
-; GFX90A-NEXT: s_mov_b64 s[0:1], 0
-; GFX90A-NEXT: s_movk_i32 s4, 0x7fff
-; GFX90A-NEXT: v_mov_b32_e32 v0, 0
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: s_and_b32 s2, s6, -4
-; GFX90A-NEXT: s_mov_b32 s3, s7
-; GFX90A-NEXT: s_load_dword s7, s[2:3], 0x0
-; GFX90A-NEXT: s_and_b32 s5, s6, 3
-; GFX90A-NEXT: s_lshl_b32 s5, s5, 3
-; GFX90A-NEXT: s_lshl_b32 s6, 0xffff, s5
-; GFX90A-NEXT: s_not_b32 s6, s6
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v1, s7
-; GFX90A-NEXT: .LBB11_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: v_mov_b32_e32 v3, v1
-; GFX90A-NEXT: v_lshrrev_b32_sdwa v1, s5, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX90A-NEXT: v_add_f32_e32 v1, 4.0, v1
-; GFX90A-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX90A-NEXT: v_add3_u32 v2, v2, v1, s4
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX90A-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc
-; GFX90A-NEXT: v_lshlrev_b32_sdwa v1, s5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; GFX90A-NEXT: v_and_or_b32 v2, v3, s6, v1
-; GFX90A-NEXT: buffer_wbl2
-; GFX90A-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[2:3] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_invl2
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v3
-; GFX90A-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX90A-NEXT: s_cbranch_execnz .LBB11_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX90A-NEXT: v_lshrrev_b32_e32 v0, s5, v1
-; GFX90A-NEXT: global_store_short v[0:1], v0, off
-; GFX90A-NEXT: s_endpgm
-;
-; GFX940-LABEL: global_atomic_fadd_ret_bf16_system:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x24
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_mov_b32_e32 v0, 0
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: s_and_b32 s2, s6, -4
-; GFX940-NEXT: s_mov_b32 s3, s7
-; GFX940-NEXT: s_load_dword s7, s[2:3], 0x0
-; GFX940-NEXT: s_and_b32 s5, s6, 3
-; GFX940-NEXT: s_lshl_b32 s5, s5, 3
-; GFX940-NEXT: s_lshl_b32 s6, 0xffff, s5
-; GFX940-NEXT: s_not_b32 s6, s6
-; GFX940-NEXT: s_waitcnt lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v1, s7
-; GFX940-NEXT: .LBB11_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: v_mov_b32_e32 v3, v1
-; GFX940-NEXT: v_lshrrev_b32_sdwa v1, s5, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_add_f32_e32 v1, 4.0, v1
-; GFX940-NEXT: v_bfe_u32 v2, v1, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX940-NEXT: v_add3_u32 v2, v2, v1, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
-; GFX940-NEXT: s_nop 1
-; GFX940-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc
-; GFX940-NEXT: v_lshlrev_b32_sdwa v1, s5, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; GFX940-NEXT: v_and_or_b32 v2, v3, s6, v1
-; GFX940-NEXT: buffer_wbl2 sc0 sc1
-; GFX940-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[2:3] sc0 sc1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc0 sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v1, v3
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB11_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: v_lshrrev_b32_e32 v0, s5, v1
-; GFX940-NEXT: global_store_short v[0:1], v0, off sc0 sc1
-; GFX940-NEXT: s_endpgm
-;
-; GFX10-LABEL: global_atomic_fadd_ret_bf16_system:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x24
-; GFX10-NEXT: v_mov_b32_e32 v0, 0
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: s_and_b32 s0, s2, -4
-; GFX10-NEXT: s_mov_b32 s1, s3
-; GFX10-NEXT: s_and_b32 s2, s2, 3
-; GFX10-NEXT: s_load_dword s3, s[0:1], 0x0
-; GFX10-NEXT: s_lshl_b32 s2, s2, 3
-; GFX10-NEXT: s_lshl_b32 s4, 0xffff, s2
-; GFX10-NEXT: s_not_b32 s4, s4
-; GFX10-NEXT: s_waitcnt lgkmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v1, s3
-; GFX10-NEXT: s_mov_b32 s3, 0
-; GFX10-NEXT: .LBB11_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: v_mov_b32_e32 v2, v1
-; GFX10-NEXT: v_lshrrev_b32_sdwa v1, s2, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX10-NEXT: v_add_f32_e32 v1, 4.0, v1
-; GFX10-NEXT: v_bfe_u32 v3, v1, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX10-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
-; GFX10-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo
-; GFX10-NEXT: v_lshlrev_b32_sdwa v1, s2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; GFX10-NEXT: v_and_or_b32 v1, v2, s4, v1
-; GFX10-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[0:1] glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
-; GFX10-NEXT: s_or_b32 s3, vcc_lo, s3
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s3
-; GFX10-NEXT: s_cbranch_execnz .LBB11_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s3
-; GFX10-NEXT: v_lshrrev_b32_e32 v0, s2, v1
-; GFX10-NEXT: global_store_short v[0:1], v0, off
-; GFX10-NEXT: s_endpgm
-;
-; GFX11-LABEL: global_atomic_fadd_ret_bf16_system:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_load_b64 s[2:3], s[0:1], 0x24
-; GFX11-NEXT: v_mov_b32_e32 v0, 0
-; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: s_and_b32 s0, s2, -4
-; GFX11-NEXT: s_mov_b32 s1, s3
-; GFX11-NEXT: s_and_b32 s2, s2, 3
-; GFX11-NEXT: s_load_b32 s3, s[0:1], 0x0
-; GFX11-NEXT: s_lshl_b32 s2, s2, 3
-; GFX11-NEXT: s_lshl_b32 s4, 0xffff, s2
-; GFX11-NEXT: s_not_b32 s4, s4
-; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v1, s3
-; GFX11-NEXT: s_mov_b32 s3, 0
-; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB11_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: v_mov_b32_e32 v2, v1
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, s2, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_add_f32_e32 v1, 4.0, v1
-; GFX11-NEXT: v_bfe_u32 v3, v1, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo
-; GFX11-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, s2, v1
-; GFX11-NEXT: v_and_or_b32 v1, v2, s4, v1
-; GFX11-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
-; GFX11-NEXT: s_or_b32 s3, vcc_lo, s3
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s3
-; GFX11-NEXT: s_cbranch_execnz .LBB11_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s3
-; GFX11-NEXT: v_lshrrev_b32_e32 v0, s2, v1
-; GFX11-NEXT: global_store_b16 v[0:1], v0, off
-; GFX11-NEXT: s_nop 0
-; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX11-NEXT: s_endpgm
-;
-; GFX12-LABEL: global_atomic_fadd_ret_bf16_system:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_load_b64 s[2:3], s[0:1], 0x24
-; GFX12-NEXT: v_mov_b32_e32 v0, 0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: s_and_b32 s0, s2, -4
-; GFX12-NEXT: s_mov_b32 s1, s3
-; GFX12-NEXT: s_and_b32 s2, s2, 3
-; GFX12-NEXT: s_load_b32 s3, s[0:1], 0x0
-; GFX12-NEXT: s_lshl_b32 s2, s2, 3
-; GFX12-NEXT: s_lshl_b32 s4, 0xffff, s2
-; GFX12-NEXT: s_not_b32 s4, s4
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v1, s3
-; GFX12-NEXT: s_mov_b32 s3, 0
-; GFX12-NEXT: .LBB11_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: v_mov_b32_e32 v2, v1
-; GFX12-NEXT: v_lshrrev_b32_e32 v1, s2, v2
-; GFX12-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX12-NEXT: v_add_f32_e32 v1, 4.0, v1
-; GFX12-NEXT: v_bfe_u32 v3, v1, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v4, 0x400000, v1
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX12-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
-; GFX12-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo
-; GFX12-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GFX12-NEXT: v_lshlrev_b32_e32 v1, s2, v1
-; GFX12-NEXT: v_and_or_b32 v1, v2, s4, v1
-; GFX12-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SYS
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2
-; GFX12-NEXT: s_or_b32 s3, vcc_lo, s3
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s3
-; GFX12-NEXT: s_cbranch_execnz .LBB11_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s3
-; GFX12-NEXT: v_lshrrev_b32_e32 v0, s2, v1
-; GFX12-NEXT: global_store_b16 v[0:1], v0, off
-; GFX12-NEXT: s_nop 0
-; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX12-NEXT: s_endpgm
- %result = atomicrmw fadd ptr addrspace(1) %ptr, bfloat 4.0 syncscope("one-as") seq_cst
- store bfloat %result, ptr addrspace(1) undef
- ret void
-}
-
-define <2 x half> @global_agent_atomic_fadd_ret_v2f16(ptr addrspace(1) %ptr, <2 x half> %val) {
-; GFX900-LABEL: global_agent_atomic_fadd_ret_v2f16:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v3, v[0:1], off
-; GFX900-NEXT: s_mov_b64 s[4:5], 0
-; GFX900-NEXT: .LBB12_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v4, v3
-; GFX900-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX900-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX900-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_cbranch_execnz .LBB12_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX900-NEXT: v_mov_b32_e32 v0, v3
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_agent_atomic_fadd_ret_v2f16:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v3, v[0:1], off
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB12_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v4, v3
-; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB12_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v0, v3
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_agent_atomic_fadd_ret_v2f16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v3, v[0:1], off
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB12_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB12_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_agent_atomic_fadd_ret_v2f16:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v3, v[0:1], off
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB12_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off sc0
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB12_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_agent_atomic_fadd_ret_v2f16:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v3, v[0:1], off
-; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: .LBB12_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v4, v3
-; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB12_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: v_mov_b32_e32 v0, v3
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_agent_atomic_fadd_ret_v2f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v3, v[0:1], off
-; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: .LBB12_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB12_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v3
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_agent_atomic_fadd_ret_v2f16:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v3, v[0:1], off
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB12_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB12_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: v_mov_b32_e32 v0, v3
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %val syncscope("agent") seq_cst
- ret <2 x half> %result
-}
-
-define <2 x half> @global_agent_atomic_fadd_ret_v2f16_offset(ptr addrspace(1) %ptr, <2 x half> %val) {
-; GFX900-LABEL: global_agent_atomic_fadd_ret_v2f16_offset:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX900-NEXT: s_mov_b64 s[4:5], 0
-; GFX900-NEXT: .LBB13_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v4, v3
-; GFX900-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX900-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:1024 glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX900-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_cbranch_execnz .LBB13_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX900-NEXT: v_mov_b32_e32 v0, v3
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_agent_atomic_fadd_ret_v2f16_offset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB13_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v4, v3
-; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:1024 glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB13_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v0, v3
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_agent_atomic_fadd_ret_v2f16_offset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB13_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:1024 glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB13_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_agent_atomic_fadd_ret_v2f16_offset:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB13_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:1024 sc0
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB13_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_agent_atomic_fadd_ret_v2f16_offset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: .LBB13_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v4, v3
-; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:1024 glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB13_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: v_mov_b32_e32 v0, v3
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_agent_atomic_fadd_ret_v2f16_offset:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:1024
-; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: .LBB13_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:1024 glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB13_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v3
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_agent_atomic_fadd_ret_v2f16_offset:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:1024
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB13_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:1024 th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB13_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: v_mov_b32_e32 v0, v3
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i32 256
- %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst
- ret <2 x half> %result
-}
-
-define <2 x half> @global_agent_atomic_fadd_ret_v2f16_negoffset(ptr addrspace(1) %ptr, <2 x half> %val) {
-; GFX900-LABEL: global_agent_atomic_fadd_ret_v2f16_negoffset:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX900-NEXT: s_mov_b64 s[4:5], 0
-; GFX900-NEXT: .LBB14_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v4, v3
-; GFX900-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX900-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-1024 glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX900-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_cbranch_execnz .LBB14_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX900-NEXT: v_mov_b32_e32 v0, v3
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_agent_atomic_fadd_ret_v2f16_negoffset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB14_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v4, v3
-; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-1024 glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB14_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v0, v3
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_agent_atomic_fadd_ret_v2f16_negoffset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB14_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-1024 glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB14_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_agent_atomic_fadd_ret_v2f16_negoffset:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB14_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-1024 sc0
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB14_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_agent_atomic_fadd_ret_v2f16_negoffset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: .LBB14_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v4, v3
-; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-1024 glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB14_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: v_mov_b32_e32 v0, v3
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_agent_atomic_fadd_ret_v2f16_negoffset:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-1024
-; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: .LBB14_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-1024 glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB14_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v3
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_agent_atomic_fadd_ret_v2f16_negoffset:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-1024
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB14_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-1024 th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB14_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: v_mov_b32_e32 v0, v3
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i32 -256
- %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst
- ret <2 x half> %result
-}
-
-define void @global_agent_atomic_fadd_noret_v2f16(ptr addrspace(1) %ptr, <2 x half> %val) {
-; GFX900-LABEL: global_agent_atomic_fadd_noret_v2f16:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v4, v[0:1], off
-; GFX900-NEXT: s_mov_b64 s[4:5], 0
-; GFX900-NEXT: .LBB15_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX900-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX900-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX900-NEXT: v_mov_b32_e32 v4, v3
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_cbranch_execnz .LBB15_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_agent_atomic_fadd_noret_v2f16:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v4, v[0:1], off
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB15_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v4, v3
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB15_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_agent_atomic_fadd_noret_v2f16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v5, v[0:1], off
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB15_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB15_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_agent_atomic_fadd_noret_v2f16:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v5, v[0:1], off
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB15_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off sc0
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB15_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_agent_atomic_fadd_noret_v2f16:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v4, v[0:1], off
-; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: .LBB15_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX10-NEXT: v_mov_b32_e32 v4, v3
-; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB15_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_agent_atomic_fadd_noret_v2f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v4, v[0:1], off
-; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: .LBB15_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX11-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB15_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_agent_atomic_fadd_noret_v2f16:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v4, v[0:1], off
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB15_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB15_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %val syncscope("agent") seq_cst
- ret void
-}
-
-define void @global_agent_atomic_fadd_noret_v2f16_offset(ptr addrspace(1) %ptr, <2 x half> %val) {
-; GFX900-LABEL: global_agent_atomic_fadd_noret_v2f16_offset:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v4, v[0:1], off offset:1024
-; GFX900-NEXT: s_mov_b64 s[4:5], 0
-; GFX900-NEXT: .LBB16_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX900-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:1024 glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX900-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX900-NEXT: v_mov_b32_e32 v4, v3
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_cbranch_execnz .LBB16_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_agent_atomic_fadd_noret_v2f16_offset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v4, v[0:1], off offset:1024
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB16_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:1024 glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v4, v3
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB16_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_agent_atomic_fadd_noret_v2f16_offset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v5, v[0:1], off offset:1024
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB16_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:1024 glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB16_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_agent_atomic_fadd_noret_v2f16_offset:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v5, v[0:1], off offset:1024
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB16_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:1024 sc0
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB16_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_agent_atomic_fadd_noret_v2f16_offset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:1024
-; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: .LBB16_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:1024 glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX10-NEXT: v_mov_b32_e32 v4, v3
-; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB16_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_agent_atomic_fadd_noret_v2f16_offset:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v4, v[0:1], off offset:1024
-; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: .LBB16_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:1024 glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX11-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB16_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_agent_atomic_fadd_noret_v2f16_offset:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v4, v[0:1], off offset:1024
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB16_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:1024 th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB16_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i32 256
- %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst
- ret void
-}
-
-define void @global_agent_atomic_fadd_noret_v2f16_negoffset(ptr addrspace(1) %ptr, <2 x half> %val) {
-; GFX900-LABEL: global_agent_atomic_fadd_noret_v2f16_negoffset:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v4, v[0:1], off offset:-1024
-; GFX900-NEXT: s_mov_b64 s[4:5], 0
-; GFX900-NEXT: .LBB17_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX900-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-1024 glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX900-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX900-NEXT: v_mov_b32_e32 v4, v3
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_cbranch_execnz .LBB17_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_agent_atomic_fadd_noret_v2f16_negoffset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v4, v[0:1], off offset:-1024
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB17_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-1024 glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v4, v3
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB17_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_agent_atomic_fadd_noret_v2f16_negoffset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v5, v[0:1], off offset:-1024
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB17_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-1024 glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB17_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_agent_atomic_fadd_noret_v2f16_negoffset:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v5, v[0:1], off offset:-1024
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB17_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-1024 sc0
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB17_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_agent_atomic_fadd_noret_v2f16_negoffset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:-1024
-; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: .LBB17_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-1024 glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX10-NEXT: v_mov_b32_e32 v4, v3
-; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB17_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_agent_atomic_fadd_noret_v2f16_negoffset:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v4, v[0:1], off offset:-1024
-; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: .LBB17_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-1024 glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX11-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB17_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_agent_atomic_fadd_noret_v2f16_negoffset:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v4, v[0:1], off offset:-1024
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB17_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-1024 th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB17_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i32 -256
- %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x half> %val syncscope("agent") seq_cst
- ret void
-}
-
-define <2 x half> @global_system_atomic_fadd_ret_v2f16(ptr addrspace(1) %ptr, <2 x half> %val) {
-; GFX900-LABEL: global_system_atomic_fadd_ret_v2f16:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v3, v[0:1], off
-; GFX900-NEXT: s_mov_b64 s[4:5], 0
-; GFX900-NEXT: .LBB18_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v4, v3
-; GFX900-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX900-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX900-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_cbranch_execnz .LBB18_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX900-NEXT: v_mov_b32_e32 v0, v3
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_system_atomic_fadd_ret_v2f16:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v3, v[0:1], off
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB18_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v4, v3
-; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB18_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v0, v3
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_system_atomic_fadd_ret_v2f16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v3, v[0:1], off
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB18_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX90A-NEXT: buffer_wbl2
-; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_invl2
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB18_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_system_atomic_fadd_ret_v2f16:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v3, v[0:1], off
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB18_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX940-NEXT: buffer_wbl2 sc0 sc1
-; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off sc0 sc1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc0 sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB18_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_system_atomic_fadd_ret_v2f16:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v3, v[0:1], off
-; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: .LBB18_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v4, v3
-; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB18_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: v_mov_b32_e32 v0, v3
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_system_atomic_fadd_ret_v2f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v3, v[0:1], off
-; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: .LBB18_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB18_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v3
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_system_atomic_fadd_ret_v2f16:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v3, v[0:1], off
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB18_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SYS
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB18_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: v_mov_b32_e32 v0, v3
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %val seq_cst
- ret <2 x half> %result
-}
-
-define <2 x half> @global_system_atomic_fadd_ret_v2f16_offset(ptr addrspace(1) %ptr, <2 x half> %val) {
-; GFX900-LABEL: global_system_atomic_fadd_ret_v2f16_offset:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX900-NEXT: s_mov_b64 s[4:5], 0
-; GFX900-NEXT: .LBB19_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v4, v3
-; GFX900-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX900-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:1024 glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX900-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_cbranch_execnz .LBB19_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX900-NEXT: v_mov_b32_e32 v0, v3
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_system_atomic_fadd_ret_v2f16_offset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB19_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v4, v3
-; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:1024 glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB19_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v0, v3
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_system_atomic_fadd_ret_v2f16_offset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB19_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX90A-NEXT: buffer_wbl2
-; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:1024 glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_invl2
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB19_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_system_atomic_fadd_ret_v2f16_offset:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB19_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX940-NEXT: buffer_wbl2 sc0 sc1
-; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:1024 sc0 sc1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc0 sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB19_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_system_atomic_fadd_ret_v2f16_offset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: .LBB19_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v4, v3
-; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:1024 glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB19_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: v_mov_b32_e32 v0, v3
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_system_atomic_fadd_ret_v2f16_offset:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:1024
-; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: .LBB19_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:1024 glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB19_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v3
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_system_atomic_fadd_ret_v2f16_offset:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:1024
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB19_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:1024 th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SYS
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB19_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: v_mov_b32_e32 v0, v3
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i32 256
- %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x half> %val seq_cst
- ret <2 x half> %result
-}
-
-define <2 x half> @global_system_atomic_fadd_ret_v2f16_negoffset(ptr addrspace(1) %ptr, <2 x half> %val) {
-; GFX900-LABEL: global_system_atomic_fadd_ret_v2f16_negoffset:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX900-NEXT: s_mov_b64 s[4:5], 0
-; GFX900-NEXT: .LBB20_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v4, v3
-; GFX900-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX900-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-1024 glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX900-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_cbranch_execnz .LBB20_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX900-NEXT: v_mov_b32_e32 v0, v3
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_system_atomic_fadd_ret_v2f16_negoffset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB20_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v4, v3
-; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-1024 glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB20_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v0, v3
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_system_atomic_fadd_ret_v2f16_negoffset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB20_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX90A-NEXT: buffer_wbl2
-; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-1024 glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_invl2
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB20_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_system_atomic_fadd_ret_v2f16_negoffset:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB20_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX940-NEXT: buffer_wbl2 sc0 sc1
-; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-1024 sc0 sc1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc0 sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB20_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_system_atomic_fadd_ret_v2f16_negoffset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: .LBB20_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v4, v3
-; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-1024 glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB20_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: v_mov_b32_e32 v0, v3
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_system_atomic_fadd_ret_v2f16_negoffset:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-1024
-; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: .LBB20_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-1024 glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB20_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v3
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_system_atomic_fadd_ret_v2f16_negoffset:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-1024
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB20_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-1024 th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SYS
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB20_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: v_mov_b32_e32 v0, v3
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i32 -256
- %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x half> %val seq_cst
- ret <2 x half> %result
-}
-
-define void @global_system_atomic_fadd_noret_v2f16(ptr addrspace(1) %ptr, <2 x half> %val) {
-; GFX900-LABEL: global_system_atomic_fadd_noret_v2f16:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v4, v[0:1], off
-; GFX900-NEXT: s_mov_b64 s[4:5], 0
-; GFX900-NEXT: .LBB21_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX900-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX900-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX900-NEXT: v_mov_b32_e32 v4, v3
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_cbranch_execnz .LBB21_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_system_atomic_fadd_noret_v2f16:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v4, v[0:1], off
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB21_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v4, v3
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB21_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_system_atomic_fadd_noret_v2f16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v5, v[0:1], off
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB21_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX90A-NEXT: buffer_wbl2
-; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_invl2
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB21_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_system_atomic_fadd_noret_v2f16:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v5, v[0:1], off
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB21_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX940-NEXT: buffer_wbl2 sc0 sc1
-; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off sc0 sc1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc0 sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB21_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_system_atomic_fadd_noret_v2f16:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v4, v[0:1], off
-; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: .LBB21_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX10-NEXT: v_mov_b32_e32 v4, v3
-; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB21_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_system_atomic_fadd_noret_v2f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v4, v[0:1], off
-; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: .LBB21_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX11-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB21_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_system_atomic_fadd_noret_v2f16:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v4, v[0:1], off
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB21_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SYS
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB21_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr addrspace(1) %ptr, <2 x half> %val seq_cst
- ret void
-}
-
-define void @global_system_atomic_fadd_noret_v2f16_offset(ptr addrspace(1) %ptr, <2 x half> %val) {
-; GFX900-LABEL: global_system_atomic_fadd_noret_v2f16_offset:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v4, v[0:1], off offset:1024
-; GFX900-NEXT: s_mov_b64 s[4:5], 0
-; GFX900-NEXT: .LBB22_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX900-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:1024 glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX900-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX900-NEXT: v_mov_b32_e32 v4, v3
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_cbranch_execnz .LBB22_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_system_atomic_fadd_noret_v2f16_offset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v4, v[0:1], off offset:1024
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB22_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:1024 glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v4, v3
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB22_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_system_atomic_fadd_noret_v2f16_offset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v5, v[0:1], off offset:1024
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB22_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX90A-NEXT: buffer_wbl2
-; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:1024 glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_invl2
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB22_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_system_atomic_fadd_noret_v2f16_offset:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v5, v[0:1], off offset:1024
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB22_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX940-NEXT: buffer_wbl2 sc0 sc1
-; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:1024 sc0 sc1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc0 sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB22_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_system_atomic_fadd_noret_v2f16_offset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:1024
-; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: .LBB22_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:1024 glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX10-NEXT: v_mov_b32_e32 v4, v3
-; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB22_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_system_atomic_fadd_noret_v2f16_offset:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v4, v[0:1], off offset:1024
-; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: .LBB22_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:1024 glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX11-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB22_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_system_atomic_fadd_noret_v2f16_offset:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v4, v[0:1], off offset:1024
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB22_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:1024 th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SYS
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB22_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i32 256
- %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x half> %val seq_cst
- ret void
-}
-
-define void @global_system_atomic_fadd_noret_v2f16_negoffset(ptr addrspace(1) %ptr, <2 x half> %val) {
-; GFX900-LABEL: global_system_atomic_fadd_noret_v2f16_negoffset:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v4, v[0:1], off offset:-1024
-; GFX900-NEXT: s_mov_b64 s[4:5], 0
-; GFX900-NEXT: .LBB23_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX900-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-1024 glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX900-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX900-NEXT: v_mov_b32_e32 v4, v3
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_cbranch_execnz .LBB23_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_system_atomic_fadd_noret_v2f16_negoffset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v4, v[0:1], off offset:-1024
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB23_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-1024 glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v4, v3
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB23_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_system_atomic_fadd_noret_v2f16_negoffset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v5, v[0:1], off offset:-1024
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB23_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX90A-NEXT: buffer_wbl2
-; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-1024 glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_invl2
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB23_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_system_atomic_fadd_noret_v2f16_negoffset:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v5, v[0:1], off offset:-1024
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB23_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX940-NEXT: buffer_wbl2 sc0 sc1
-; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[4:5], off offset:-1024 sc0 sc1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc0 sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB23_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_system_atomic_fadd_noret_v2f16_negoffset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v4, v[0:1], off offset:-1024
-; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: .LBB23_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off offset:-1024 glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX10-NEXT: v_mov_b32_e32 v4, v3
-; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB23_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_system_atomic_fadd_noret_v2f16_negoffset:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v4, v[0:1], off offset:-1024
-; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: .LBB23_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-1024 glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX11-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB23_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_system_atomic_fadd_noret_v2f16_negoffset:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v4, v[0:1], off offset:-1024
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB23_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:-1024 th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SYS
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB23_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x half>, ptr addrspace(1) %ptr, i32 -256
- %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x half> %val seq_cst
- ret void
-}
-
-define <2 x bfloat> @global_agent_atomic_fadd_ret_v2bf16(ptr addrspace(1) %ptr, <2 x bfloat> %val) {
-; GFX900-LABEL: global_agent_atomic_fadd_ret_v2bf16:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v3, v[0:1], off
-; GFX900-NEXT: s_mov_b64 s[6:7], 0
-; GFX900-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX900-NEXT: s_movk_i32 s8, 0x7fff
-; GFX900-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX900-NEXT: s_mov_b32 s9, 0x7060302
-; GFX900-NEXT: .LBB24_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v6, v3
-; GFX900-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX900-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX900-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX900-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX900-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX900-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX900-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX900-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX900-NEXT: v_add3_u32 v7, v7, v3, s8
-; GFX900-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX900-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX900-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX900-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
-; GFX900-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX900-NEXT: v_perm_b32 v5, v5, v3, s9
-; GFX900-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
-; GFX900-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX900-NEXT: s_cbranch_execnz .LBB24_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX900-NEXT: v_mov_b32_e32 v0, v3
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_agent_atomic_fadd_ret_v2bf16:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v3, v[0:1], off
-; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB24_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v6, v3
-; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX908-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
-; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
-; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
-; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB24_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: v_mov_b32_e32 v0, v3
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_agent_atomic_fadd_ret_v2bf16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v3, v[0:1], off
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB24_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v7, v3
-; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
-; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
-; GFX90A-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
-; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
-; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB24_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_agent_atomic_fadd_ret_v2bf16:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v3, v[0:1], off
-; GFX940-NEXT: s_mov_b64 s[2:3], 0
-; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB24_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v7, v3
-; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
-; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
-; GFX940-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
-; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
-; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
-; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off sc0
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
-; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB24_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_agent_atomic_fadd_ret_v2bf16:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v3, v[0:1], off
-; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX10-NEXT: s_mov_b32 s5, 0
-; GFX10-NEXT: .LBB24_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v6, v3
-; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX10-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
-; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
-; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_cbranch_execnz .LBB24_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: v_mov_b32_e32 v0, v3
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_agent_atomic_fadd_ret_v2bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v3, v[0:1], off
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: s_mov_b32 s1, 0
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB24_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v6, v3
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
-; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_cbranch_execnz .LBB24_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: v_mov_b32_e32 v0, v3
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_agent_atomic_fadd_ret_v2bf16:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v3, v[0:1], off
-; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB24_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v6, v3
-; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX12-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
-; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
-; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB24_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: v_mov_b32_e32 v0, v3
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr addrspace(1) %ptr, <2 x bfloat> %val syncscope("agent") seq_cst
- ret <2 x bfloat> %result
-}
-
-define <2 x bfloat> @global_agent_atomic_fadd_ret_v2bf16_offset(ptr addrspace(1) %ptr, <2 x bfloat> %val) {
-; GFX900-LABEL: global_agent_atomic_fadd_ret_v2bf16_offset:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX900-NEXT: s_mov_b64 s[6:7], 0
-; GFX900-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX900-NEXT: s_movk_i32 s8, 0x7fff
-; GFX900-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX900-NEXT: s_mov_b32 s9, 0x7060302
-; GFX900-NEXT: .LBB25_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v6, v3
-; GFX900-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX900-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX900-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX900-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX900-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX900-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX900-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX900-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX900-NEXT: v_add3_u32 v7, v7, v3, s8
-; GFX900-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX900-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX900-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX900-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
-; GFX900-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX900-NEXT: v_perm_b32 v5, v5, v3, s9
-; GFX900-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:1024 glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
-; GFX900-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX900-NEXT: s_cbranch_execnz .LBB25_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX900-NEXT: v_mov_b32_e32 v0, v3
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_agent_atomic_fadd_ret_v2bf16_offset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB25_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v6, v3
-; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX908-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
-; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
-; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:1024 glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
-; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB25_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: v_mov_b32_e32 v0, v3
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_agent_atomic_fadd_ret_v2bf16_offset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB25_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v7, v3
-; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
-; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
-; GFX90A-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
-; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
-; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:1024 glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB25_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_agent_atomic_fadd_ret_v2bf16_offset:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX940-NEXT: s_mov_b64 s[2:3], 0
-; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB25_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v7, v3
-; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
-; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
-; GFX940-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
-; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
-; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
-; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:1024 sc0
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
-; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB25_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_agent_atomic_fadd_ret_v2bf16_offset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX10-NEXT: s_mov_b32 s5, 0
-; GFX10-NEXT: .LBB25_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v6, v3
-; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX10-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
-; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
-; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:1024 glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_cbranch_execnz .LBB25_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: v_mov_b32_e32 v0, v3
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_agent_atomic_fadd_ret_v2bf16_offset:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:1024
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: s_mov_b32 s1, 0
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB25_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v6, v3
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
-; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:1024 glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_cbranch_execnz .LBB25_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: v_mov_b32_e32 v0, v3
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_agent_atomic_fadd_ret_v2bf16_offset:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:1024
-; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB25_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v6, v3
-; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX12-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
-; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
-; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:1024 th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB25_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: v_mov_b32_e32 v0, v3
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i32 256
- %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
- ret <2 x bfloat> %result
-}
-
-define <2 x bfloat> @global_agent_atomic_fadd_ret_v2bf16_negoffset(ptr addrspace(1) %ptr, <2 x bfloat> %val) {
-; GFX900-LABEL: global_agent_atomic_fadd_ret_v2bf16_negoffset:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX900-NEXT: s_mov_b64 s[6:7], 0
-; GFX900-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX900-NEXT: s_movk_i32 s8, 0x7fff
-; GFX900-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX900-NEXT: s_mov_b32 s9, 0x7060302
-; GFX900-NEXT: .LBB26_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v6, v3
-; GFX900-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX900-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX900-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX900-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX900-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX900-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX900-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX900-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX900-NEXT: v_add3_u32 v7, v7, v3, s8
-; GFX900-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX900-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX900-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX900-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
-; GFX900-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX900-NEXT: v_perm_b32 v5, v5, v3, s9
-; GFX900-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:-1024 glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
-; GFX900-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX900-NEXT: s_cbranch_execnz .LBB26_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX900-NEXT: v_mov_b32_e32 v0, v3
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_agent_atomic_fadd_ret_v2bf16_negoffset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB26_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v6, v3
-; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX908-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
-; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
-; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:-1024 glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
-; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB26_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: v_mov_b32_e32 v0, v3
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_agent_atomic_fadd_ret_v2bf16_negoffset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB26_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v7, v3
-; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
-; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
-; GFX90A-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
-; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
-; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:-1024 glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB26_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_agent_atomic_fadd_ret_v2bf16_negoffset:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX940-NEXT: s_mov_b64 s[2:3], 0
-; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB26_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v7, v3
-; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
-; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
-; GFX940-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
-; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
-; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
-; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:-1024 sc0
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
-; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB26_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_agent_atomic_fadd_ret_v2bf16_negoffset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX10-NEXT: s_mov_b32 s5, 0
-; GFX10-NEXT: .LBB26_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v6, v3
-; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX10-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
-; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
-; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:-1024 glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_cbranch_execnz .LBB26_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: v_mov_b32_e32 v0, v3
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_agent_atomic_fadd_ret_v2bf16_negoffset:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-1024
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: s_mov_b32 s1, 0
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB26_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v6, v3
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
-; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:-1024 glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_cbranch_execnz .LBB26_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: v_mov_b32_e32 v0, v3
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_agent_atomic_fadd_ret_v2bf16_negoffset:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-1024
-; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB26_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v6, v3
-; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX12-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
-; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
-; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:-1024 th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB26_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: v_mov_b32_e32 v0, v3
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i32 -256
- %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
- ret <2 x bfloat> %result
-}
-
-define void @global_agent_atomic_fadd_noret_v2bf16(ptr addrspace(1) %ptr, <2 x bfloat> %val) {
-; GFX900-LABEL: global_agent_atomic_fadd_noret_v2bf16:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v3, v[0:1], off
-; GFX900-NEXT: s_mov_b64 s[6:7], 0
-; GFX900-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX900-NEXT: s_movk_i32 s8, 0x7fff
-; GFX900-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX900-NEXT: s_mov_b32 s9, 0x7060302
-; GFX900-NEXT: .LBB27_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX900-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX900-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX900-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX900-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX900-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX900-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX900-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX900-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX900-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX900-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX900-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX900-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX900-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX900-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX900-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX900-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX900-NEXT: v_mov_b32_e32 v3, v2
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX900-NEXT: s_cbranch_execnz .LBB27_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_agent_atomic_fadd_noret_v2bf16:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v3, v[0:1], off
-; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB27_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX908-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX908-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX908-NEXT: v_mov_b32_e32 v3, v2
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB27_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_agent_atomic_fadd_noret_v2bf16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v3, v[0:1], off
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB27_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX90A-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX90A-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: v_mov_b32_e32 v3, v2
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB27_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_agent_atomic_fadd_noret_v2bf16:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v3, v[0:1], off
-; GFX940-NEXT: s_mov_b64 s[2:3], 0
-; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB27_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX940-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX940-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
-; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
-; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB27_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_agent_atomic_fadd_noret_v2bf16:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v3, v[0:1], off
-; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX10-NEXT: s_mov_b32 s5, 0
-; GFX10-NEXT: .LBB27_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX10-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX10-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
-; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
-; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX10-NEXT: v_mov_b32_e32 v3, v2
-; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_cbranch_execnz .LBB27_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_agent_atomic_fadd_noret_v2bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v3, v[0:1], off
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX11-NEXT: s_mov_b32 s1, 0
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB27_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX11-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX11-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
-; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX11-NEXT: v_mov_b32_e32 v3, v2
-; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_cbranch_execnz .LBB27_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_agent_atomic_fadd_noret_v2bf16:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v3, v[0:1], off
-; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB27_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX12-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX12-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
-; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
-; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX12-NEXT: v_mov_b32_e32 v3, v2
-; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB27_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr addrspace(1) %ptr, <2 x bfloat> %val syncscope("agent") seq_cst
- ret void
-}
-
-define void @global_agent_atomic_fadd_noret_v2bf16_offset(ptr addrspace(1) %ptr, <2 x bfloat> %val) {
-; GFX900-LABEL: global_agent_atomic_fadd_noret_v2bf16_offset:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX900-NEXT: s_mov_b64 s[6:7], 0
-; GFX900-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX900-NEXT: s_movk_i32 s8, 0x7fff
-; GFX900-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX900-NEXT: s_mov_b32 s9, 0x7060302
-; GFX900-NEXT: .LBB28_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX900-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX900-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX900-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX900-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX900-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX900-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX900-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX900-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX900-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX900-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX900-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX900-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX900-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX900-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX900-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:1024 glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX900-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX900-NEXT: v_mov_b32_e32 v3, v2
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX900-NEXT: s_cbranch_execnz .LBB28_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_agent_atomic_fadd_noret_v2bf16_offset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB28_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX908-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX908-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:1024 glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX908-NEXT: v_mov_b32_e32 v3, v2
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB28_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_agent_atomic_fadd_noret_v2bf16_offset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB28_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX90A-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX90A-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:1024 glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: v_mov_b32_e32 v3, v2
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB28_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_agent_atomic_fadd_noret_v2bf16_offset:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX940-NEXT: s_mov_b64 s[2:3], 0
-; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB28_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX940-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX940-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
-; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
-; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:1024 sc0
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB28_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_agent_atomic_fadd_noret_v2bf16_offset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX10-NEXT: s_mov_b32 s5, 0
-; GFX10-NEXT: .LBB28_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX10-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX10-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
-; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
-; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:1024 glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX10-NEXT: v_mov_b32_e32 v3, v2
-; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_cbranch_execnz .LBB28_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_agent_atomic_fadd_noret_v2bf16_offset:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:1024
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX11-NEXT: s_mov_b32 s1, 0
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB28_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX11-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX11-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
-; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:1024 glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX11-NEXT: v_mov_b32_e32 v3, v2
-; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_cbranch_execnz .LBB28_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_agent_atomic_fadd_noret_v2bf16_offset:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:1024
-; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB28_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX12-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX12-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
-; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
-; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:1024 th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX12-NEXT: v_mov_b32_e32 v3, v2
-; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB28_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i32 256
- %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
- ret void
-}
-
-define void @global_agent_atomic_fadd_noret_v2bf16_negoffset(ptr addrspace(1) %ptr, <2 x bfloat> %val) {
-; GFX900-LABEL: global_agent_atomic_fadd_noret_v2bf16_negoffset:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX900-NEXT: s_mov_b64 s[6:7], 0
-; GFX900-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX900-NEXT: s_movk_i32 s8, 0x7fff
-; GFX900-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX900-NEXT: s_mov_b32 s9, 0x7060302
-; GFX900-NEXT: .LBB29_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX900-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX900-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX900-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX900-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX900-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX900-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX900-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX900-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX900-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX900-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX900-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX900-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX900-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX900-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX900-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-1024 glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX900-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX900-NEXT: v_mov_b32_e32 v3, v2
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX900-NEXT: s_cbranch_execnz .LBB29_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_agent_atomic_fadd_noret_v2bf16_negoffset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB29_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX908-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX908-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-1024 glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX908-NEXT: v_mov_b32_e32 v3, v2
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB29_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_agent_atomic_fadd_noret_v2bf16_negoffset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB29_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX90A-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX90A-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-1024 glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: v_mov_b32_e32 v3, v2
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB29_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_agent_atomic_fadd_noret_v2bf16_negoffset:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX940-NEXT: s_mov_b64 s[2:3], 0
-; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB29_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX940-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX940-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
-; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
-; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-1024 sc0
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB29_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_agent_atomic_fadd_noret_v2bf16_negoffset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX10-NEXT: s_mov_b32 s5, 0
-; GFX10-NEXT: .LBB29_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX10-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX10-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
-; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
-; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-1024 glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX10-NEXT: v_mov_b32_e32 v3, v2
-; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_cbranch_execnz .LBB29_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_agent_atomic_fadd_noret_v2bf16_negoffset:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-1024
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX11-NEXT: s_mov_b32 s1, 0
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB29_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX11-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX11-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
-; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-1024 glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX11-NEXT: v_mov_b32_e32 v3, v2
-; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_cbranch_execnz .LBB29_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_agent_atomic_fadd_noret_v2bf16_negoffset:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-1024
-; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB29_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX12-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX12-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
-; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
-; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-1024 th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX12-NEXT: v_mov_b32_e32 v3, v2
-; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB29_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i32 -256
- %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
- ret void
-}
-
-define <2 x bfloat> @global_system_atomic_fadd_ret_v2bf16(ptr addrspace(1) %ptr, <2 x bfloat> %val) {
-; GFX900-LABEL: global_system_atomic_fadd_ret_v2bf16:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v3, v[0:1], off
-; GFX900-NEXT: s_mov_b64 s[6:7], 0
-; GFX900-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX900-NEXT: s_movk_i32 s8, 0x7fff
-; GFX900-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX900-NEXT: s_mov_b32 s9, 0x7060302
-; GFX900-NEXT: .LBB30_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v6, v3
-; GFX900-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX900-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX900-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX900-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX900-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX900-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX900-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX900-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX900-NEXT: v_add3_u32 v7, v7, v3, s8
-; GFX900-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX900-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX900-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX900-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
-; GFX900-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX900-NEXT: v_perm_b32 v5, v5, v3, s9
-; GFX900-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
-; GFX900-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX900-NEXT: s_cbranch_execnz .LBB30_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX900-NEXT: v_mov_b32_e32 v0, v3
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_system_atomic_fadd_ret_v2bf16:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v3, v[0:1], off
-; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB30_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v6, v3
-; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX908-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
-; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
-; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
-; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB30_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: v_mov_b32_e32 v0, v3
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_system_atomic_fadd_ret_v2bf16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v3, v[0:1], off
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB30_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v7, v3
-; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
-; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
-; GFX90A-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
-; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
-; GFX90A-NEXT: buffer_wbl2
-; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_invl2
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB30_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_system_atomic_fadd_ret_v2bf16:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v3, v[0:1], off
-; GFX940-NEXT: s_mov_b64 s[2:3], 0
-; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB30_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v7, v3
-; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
-; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
-; GFX940-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
-; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
-; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
-; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
-; GFX940-NEXT: buffer_wbl2 sc0 sc1
-; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off sc0 sc1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc0 sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
-; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB30_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_system_atomic_fadd_ret_v2bf16:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v3, v[0:1], off
-; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX10-NEXT: s_mov_b32 s5, 0
-; GFX10-NEXT: .LBB30_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v6, v3
-; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX10-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
-; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
-; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_cbranch_execnz .LBB30_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: v_mov_b32_e32 v0, v3
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_system_atomic_fadd_ret_v2bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v3, v[0:1], off
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: s_mov_b32 s1, 0
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB30_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v6, v3
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
-; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_cbranch_execnz .LBB30_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: v_mov_b32_e32 v0, v3
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_system_atomic_fadd_ret_v2bf16:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v3, v[0:1], off
-; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB30_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v6, v3
-; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX12-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
-; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
-; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SYS
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB30_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: v_mov_b32_e32 v0, v3
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr addrspace(1) %ptr, <2 x bfloat> %val seq_cst
- ret <2 x bfloat> %result
-}
-
-define <2 x bfloat> @global_system_atomic_fadd_ret_v2bf16_offset(ptr addrspace(1) %ptr, <2 x bfloat> %val) {
-; GFX900-LABEL: global_system_atomic_fadd_ret_v2bf16_offset:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX900-NEXT: s_mov_b64 s[6:7], 0
-; GFX900-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX900-NEXT: s_movk_i32 s8, 0x7fff
-; GFX900-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX900-NEXT: s_mov_b32 s9, 0x7060302
-; GFX900-NEXT: .LBB31_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v6, v3
-; GFX900-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX900-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX900-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX900-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX900-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX900-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX900-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX900-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX900-NEXT: v_add3_u32 v7, v7, v3, s8
-; GFX900-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX900-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX900-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX900-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
-; GFX900-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX900-NEXT: v_perm_b32 v5, v5, v3, s9
-; GFX900-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:1024 glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
-; GFX900-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX900-NEXT: s_cbranch_execnz .LBB31_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX900-NEXT: v_mov_b32_e32 v0, v3
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_system_atomic_fadd_ret_v2bf16_offset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB31_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v6, v3
-; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX908-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
-; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
-; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:1024 glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
-; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB31_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: v_mov_b32_e32 v0, v3
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_system_atomic_fadd_ret_v2bf16_offset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB31_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v7, v3
-; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
-; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
-; GFX90A-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
-; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
-; GFX90A-NEXT: buffer_wbl2
-; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:1024 glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_invl2
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB31_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_system_atomic_fadd_ret_v2bf16_offset:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX940-NEXT: s_mov_b64 s[2:3], 0
-; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB31_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v7, v3
-; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
-; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
-; GFX940-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
-; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
-; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
-; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
-; GFX940-NEXT: buffer_wbl2 sc0 sc1
-; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:1024 sc0 sc1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc0 sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
-; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB31_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_system_atomic_fadd_ret_v2bf16_offset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX10-NEXT: s_mov_b32 s5, 0
-; GFX10-NEXT: .LBB31_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v6, v3
-; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX10-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
-; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
-; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:1024 glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_cbranch_execnz .LBB31_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: v_mov_b32_e32 v0, v3
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_system_atomic_fadd_ret_v2bf16_offset:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:1024
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: s_mov_b32 s1, 0
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB31_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v6, v3
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
-; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:1024 glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_cbranch_execnz .LBB31_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: v_mov_b32_e32 v0, v3
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_system_atomic_fadd_ret_v2bf16_offset:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:1024
-; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB31_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v6, v3
-; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX12-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
-; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
-; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:1024 th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SYS
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB31_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: v_mov_b32_e32 v0, v3
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i32 256
- %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x bfloat> %val seq_cst
- ret <2 x bfloat> %result
-}
-
-define <2 x bfloat> @global_system_atomic_fadd_ret_v2bf16_negoffset(ptr addrspace(1) %ptr, <2 x bfloat> %val) {
-; GFX900-LABEL: global_system_atomic_fadd_ret_v2bf16_negoffset:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX900-NEXT: s_mov_b64 s[6:7], 0
-; GFX900-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX900-NEXT: s_movk_i32 s8, 0x7fff
-; GFX900-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX900-NEXT: s_mov_b32 s9, 0x7060302
-; GFX900-NEXT: .LBB32_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_mov_b32_e32 v6, v3
-; GFX900-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX900-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX900-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX900-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX900-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX900-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX900-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX900-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX900-NEXT: v_add3_u32 v7, v7, v3, s8
-; GFX900-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX900-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX900-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX900-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
-; GFX900-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX900-NEXT: v_perm_b32 v5, v5, v3, s9
-; GFX900-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:-1024 glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
-; GFX900-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX900-NEXT: s_cbranch_execnz .LBB32_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX900-NEXT: v_mov_b32_e32 v0, v3
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_system_atomic_fadd_ret_v2bf16_negoffset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB32_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v6, v3
-; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX908-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
-; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
-; GFX908-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:-1024 glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
-; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB32_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: v_mov_b32_e32 v0, v3
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_system_atomic_fadd_ret_v2bf16_negoffset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB32_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v7, v3
-; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
-; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
-; GFX90A-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
-; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
-; GFX90A-NEXT: buffer_wbl2
-; GFX90A-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:-1024 glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_invl2
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB32_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_system_atomic_fadd_ret_v2bf16_negoffset:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX940-NEXT: s_mov_b64 s[2:3], 0
-; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB32_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v7, v3
-; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
-; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
-; GFX940-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
-; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
-; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
-; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
-; GFX940-NEXT: buffer_wbl2 sc0 sc1
-; GFX940-NEXT: global_atomic_cmpswap v3, v[0:1], v[6:7], off offset:-1024 sc0 sc1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc0 sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
-; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB32_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_system_atomic_fadd_ret_v2bf16_negoffset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX10-NEXT: s_mov_b32 s5, 0
-; GFX10-NEXT: .LBB32_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v6, v3
-; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX10-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
-; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
-; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v3, v[0:1], v[5:6], off offset:-1024 glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_cbranch_execnz .LBB32_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: v_mov_b32_e32 v0, v3
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_system_atomic_fadd_ret_v2bf16_negoffset:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-1024
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: s_mov_b32 s1, 0
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB32_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v6, v3
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
-; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:-1024 glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_cbranch_execnz .LBB32_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: v_mov_b32_e32 v0, v3
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_system_atomic_fadd_ret_v2bf16_negoffset:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-1024
-; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB32_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v6, v3
-; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX12-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
-; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
-; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[5:6], off offset:-1024 th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SYS
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB32_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: v_mov_b32_e32 v0, v3
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i32 -256
- %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x bfloat> %val seq_cst
- ret <2 x bfloat> %result
-}
-
-define void @global_system_atomic_fadd_noret_v2bf16(ptr addrspace(1) %ptr, <2 x bfloat> %val) {
-; GFX900-LABEL: global_system_atomic_fadd_noret_v2bf16:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v3, v[0:1], off
-; GFX900-NEXT: s_mov_b64 s[6:7], 0
-; GFX900-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX900-NEXT: s_movk_i32 s8, 0x7fff
-; GFX900-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX900-NEXT: s_mov_b32 s9, 0x7060302
-; GFX900-NEXT: .LBB33_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX900-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX900-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX900-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX900-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX900-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX900-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX900-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX900-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX900-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX900-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX900-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX900-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX900-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX900-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX900-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX900-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX900-NEXT: v_mov_b32_e32 v3, v2
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX900-NEXT: s_cbranch_execnz .LBB33_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_system_atomic_fadd_noret_v2bf16:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v3, v[0:1], off
-; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB33_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX908-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX908-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX908-NEXT: v_mov_b32_e32 v3, v2
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB33_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_system_atomic_fadd_noret_v2bf16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v3, v[0:1], off
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB33_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX90A-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX90A-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX90A-NEXT: buffer_wbl2
-; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_invl2
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: v_mov_b32_e32 v3, v2
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB33_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_system_atomic_fadd_noret_v2bf16:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v3, v[0:1], off
-; GFX940-NEXT: s_mov_b64 s[2:3], 0
-; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB33_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX940-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX940-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
-; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
-; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
-; GFX940-NEXT: buffer_wbl2 sc0 sc1
-; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off sc0 sc1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc0 sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB33_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_system_atomic_fadd_noret_v2bf16:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v3, v[0:1], off
-; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX10-NEXT: s_mov_b32 s5, 0
-; GFX10-NEXT: .LBB33_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX10-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX10-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
-; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
-; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX10-NEXT: v_mov_b32_e32 v3, v2
-; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_cbranch_execnz .LBB33_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_system_atomic_fadd_noret_v2bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v3, v[0:1], off
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX11-NEXT: s_mov_b32 s1, 0
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB33_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX11-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX11-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
-; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX11-NEXT: v_mov_b32_e32 v3, v2
-; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_cbranch_execnz .LBB33_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_system_atomic_fadd_noret_v2bf16:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v3, v[0:1], off
-; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB33_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX12-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX12-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
-; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
-; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SYS
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX12-NEXT: v_mov_b32_e32 v3, v2
-; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB33_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr addrspace(1) %ptr, <2 x bfloat> %val seq_cst
- ret void
-}
-
-define void @global_system_atomic_fadd_noret_v2bf16_offset(ptr addrspace(1) %ptr, <2 x bfloat> %val) {
-; GFX900-LABEL: global_system_atomic_fadd_noret_v2bf16_offset:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX900-NEXT: s_mov_b64 s[6:7], 0
-; GFX900-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX900-NEXT: s_movk_i32 s8, 0x7fff
-; GFX900-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX900-NEXT: s_mov_b32 s9, 0x7060302
-; GFX900-NEXT: .LBB34_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX900-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX900-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX900-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX900-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX900-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX900-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX900-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX900-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX900-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX900-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX900-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX900-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX900-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX900-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX900-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:1024 glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX900-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX900-NEXT: v_mov_b32_e32 v3, v2
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX900-NEXT: s_cbranch_execnz .LBB34_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_system_atomic_fadd_noret_v2bf16_offset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB34_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX908-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX908-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:1024 glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX908-NEXT: v_mov_b32_e32 v3, v2
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB34_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_system_atomic_fadd_noret_v2bf16_offset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB34_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX90A-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX90A-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX90A-NEXT: buffer_wbl2
-; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:1024 glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_invl2
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: v_mov_b32_e32 v3, v2
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB34_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_system_atomic_fadd_noret_v2bf16_offset:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX940-NEXT: s_mov_b64 s[2:3], 0
-; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB34_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX940-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX940-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
-; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
-; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
-; GFX940-NEXT: buffer_wbl2 sc0 sc1
-; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:1024 sc0 sc1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc0 sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB34_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_system_atomic_fadd_noret_v2bf16_offset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:1024
-; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX10-NEXT: s_mov_b32 s5, 0
-; GFX10-NEXT: .LBB34_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX10-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX10-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
-; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
-; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:1024 glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX10-NEXT: v_mov_b32_e32 v3, v2
-; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_cbranch_execnz .LBB34_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_system_atomic_fadd_noret_v2bf16_offset:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:1024
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX11-NEXT: s_mov_b32 s1, 0
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB34_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX11-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX11-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
-; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:1024 glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX11-NEXT: v_mov_b32_e32 v3, v2
-; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_cbranch_execnz .LBB34_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_system_atomic_fadd_noret_v2bf16_offset:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:1024
-; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB34_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX12-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX12-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
-; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
-; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:1024 th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SYS
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX12-NEXT: v_mov_b32_e32 v3, v2
-; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB34_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i32 256
- %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x bfloat> %val seq_cst
- ret void
-}
-
-define void @global_system_atomic_fadd_noret_v2bf16_negoffset(ptr addrspace(1) %ptr, <2 x bfloat> %val) {
-; GFX900-LABEL: global_system_atomic_fadd_noret_v2bf16_negoffset:
-; GFX900: ; %bb.0:
-; GFX900-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX900-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX900-NEXT: s_mov_b64 s[6:7], 0
-; GFX900-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX900-NEXT: s_movk_i32 s8, 0x7fff
-; GFX900-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX900-NEXT: s_mov_b32 s9, 0x7060302
-; GFX900-NEXT: .LBB35_1: ; %atomicrmw.start
-; GFX900-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX900-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX900-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX900-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX900-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX900-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX900-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX900-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX900-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX900-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX900-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX900-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX900-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX900-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX900-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX900-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-1024 glc
-; GFX900-NEXT: s_waitcnt vmcnt(0)
-; GFX900-NEXT: buffer_wbinvl1_vol
-; GFX900-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX900-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX900-NEXT: v_mov_b32_e32 v3, v2
-; GFX900-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX900-NEXT: s_cbranch_execnz .LBB35_1
-; GFX900-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX900-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX900-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: global_system_atomic_fadd_noret_v2bf16_negoffset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB35_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX908-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX908-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX908-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-1024 glc
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX908-NEXT: v_mov_b32_e32 v3, v2
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB35_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: global_system_atomic_fadd_noret_v2bf16_negoffset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB35_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX90A-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX90A-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX90A-NEXT: buffer_wbl2
-; GFX90A-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-1024 glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: buffer_invl2
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: v_mov_b32_e32 v3, v2
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB35_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: global_system_atomic_fadd_noret_v2bf16_negoffset:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX940-NEXT: s_mov_b64 s[2:3], 0
-; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB35_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX940-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX940-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
-; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
-; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
-; GFX940-NEXT: buffer_wbl2 sc0 sc1
-; GFX940-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-1024 sc0 sc1
-; GFX940-NEXT: s_waitcnt vmcnt(0)
-; GFX940-NEXT: buffer_inv sc0 sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB35_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: global_system_atomic_fadd_noret_v2bf16_negoffset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: global_load_dword v3, v[0:1], off offset:-1024
-; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX10-NEXT: s_mov_b32 s5, 0
-; GFX10-NEXT: .LBB35_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX10-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX10-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
-; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
-; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: global_atomic_cmpswap v2, v[0:1], v[2:3], off offset:-1024 glc
-; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX10-NEXT: v_mov_b32_e32 v3, v2
-; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_cbranch_execnz .LBB35_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: global_system_atomic_fadd_noret_v2bf16_negoffset:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: global_load_b32 v3, v[0:1], off offset:-1024
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX11-NEXT: s_mov_b32 s1, 0
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB35_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX11-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX11-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
-; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-1024 glc
-; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX11-NEXT: v_mov_b32_e32 v3, v2
-; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_cbranch_execnz .LBB35_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: global_system_atomic_fadd_noret_v2bf16_negoffset:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: global_load_b32 v3, v[0:1], off offset:-1024
-; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB35_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX12-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX12-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
-; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
-; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:-1024 th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SYS
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX12-NEXT: v_mov_b32_e32 v3, v2
-; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB35_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x bfloat>, ptr addrspace(1) %ptr, i32 -256
- %result = atomicrmw fadd ptr addrspace(1) %gep, <2 x bfloat> %val seq_cst
- ret void
-}
-
-attributes #0 = { "denormal-fp-math-f32"="preserve-sign,preserve-sign" "amdgpu-unsafe-fp-atomics"="true" }
-attributes #1 = { "denormal-fp-math-f32"="preserve-sign,preserve-sign" "target-cpu"="gfx803" "target-features"="+atomic-fadd-no-rtn-insts" "amdgpu-unsafe-fp-atomics"="true" }
-attributes #2 = { "amdgpu-unsafe-fp-atomics"="true" }
>From 3908103758fde5c62d255c4d36eb57d4dc78831f Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Tue, 11 Jun 2024 16:52:08 +0200
Subject: [PATCH 3/4] AMDGPU: Add flat atomicrmw tests for fadd/fmin/fmax/fsub
Make flat tests consistent with other address spaces.
---
.../CodeGen/AMDGPU/flat-atomicrmw-fadd.ll | 17959 ++++++++++++----
.../CodeGen/AMDGPU/flat-atomicrmw-fmax.ll | 15756 ++++++++++++++
.../CodeGen/AMDGPU/flat-atomicrmw-fmin.ll | 15756 ++++++++++++++
.../CodeGen/AMDGPU/flat-atomicrmw-fsub.ll | 15075 +++++++++++++
4 files changed, 60441 insertions(+), 4105 deletions(-)
create mode 100644 llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmax.ll
create mode 100644 llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmin.ll
create mode 100644 llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fsub.ll
diff --git a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fadd.ll b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fadd.ll
index 1e7cf0e702a03..cdfb71b9bf6b6 100644
--- a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fadd.ll
+++ b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fadd.ll
@@ -1,82 +1,125 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 < %s | FileCheck -check-prefix=GFX7 %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx803 < %s | FileCheck -check-prefix=GFX8 %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 < %s | FileCheck -check-prefixes=GFX90,GFX908 %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a < %s | FileCheck -check-prefixes=GFX90,GFX90A %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx940 < %s | FileCheck -check-prefix=GFX940 %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11 %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 < %s | FileCheck -check-prefix=GFX12 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 < %s | FileCheck -check-prefix=GFX12 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx940 < %s | FileCheck -check-prefix=GFX940 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx90a < %s | FileCheck -check-prefix=GFX90A %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx908 < %s | FileCheck -check-prefix=GFX908 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tonga < %s | FileCheck -check-prefix=GFX8 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=hawaii < %s | FileCheck -check-prefix=GFX7 %s
-define <2 x half> @flat_agent_atomic_fadd_ret_v2f16(ptr %ptr, <2 x half> %val) {
-; GFX7-LABEL: flat_agent_atomic_fadd_ret_v2f16:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: flat_load_dword v5, v[0:1]
-; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v3
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v6
-; GFX7-NEXT: .LBB0_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
-; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX7-NEXT: v_add_f32_e32 v6, v6, v4
-; GFX7-NEXT: v_add_f32_e32 v7, v7, v5
-; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
-; GFX7-NEXT: v_or_b32_e32 v7, v2, v3
-; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v6
-; GFX7-NEXT: v_or_b32_e32 v6, v8, v2
-; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[6:7] glc
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: buffer_wbinvl1_vol
-; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v6
-; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v6
-; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v7
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB0_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: v_mov_b32_e32 v0, v2
-; GFX7-NEXT: v_mov_b32_e32 v1, v3
-; GFX7-NEXT: s_setpc_b64 s[30:31]
+; --------------------------------------------------------------------
+; float
+; --------------------------------------------------------------------
+
+define float @flat_agent_atomic_fadd_ret_f32(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
;
-; GFX8-LABEL: flat_agent_atomic_fadd_ret_v2f16:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: flat_load_dword v3, v[0:1]
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB0_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_mov_b32_e32 v4, v3
-; GFX8-NEXT: v_add_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
-; GFX8-NEXT: v_add_f16_e32 v5, v4, v2
-; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
-; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: buffer_wbinvl1_vol
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB0_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v0, v3
-; GFX8-NEXT: s_setpc_b64 s[30:31]
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX908-LABEL: flat_agent_atomic_fadd_ret_v2f16:
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB0_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: s_mov_b64 s[4:5], src_shared_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s5, v1
+; GFX90A-NEXT: ; implicit-def: $vgpr3
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX90A-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execz .LBB0_6
+; GFX90A-NEXT: ; %bb.1: ; %atomicrmw.check.private
+; GFX90A-NEXT: s_mov_b64 s[6:7], src_private_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s7, v1
+; GFX90A-NEXT: ; implicit-def: $vgpr3
+; GFX90A-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX90A-NEXT: s_xor_b64 s[6:7], exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB0_3
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.global
+; GFX90A-NEXT: global_atomic_add_f32 v3, v[0:1], v2, off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: .LBB0_3: ; %Flow
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[6:7], s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB0_5
+; GFX90A-NEXT: ; %bb.4: ; %atomicrmw.private
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX90A-NEXT: buffer_load_dword v3, v0, s[0:3], 0 offen
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f32_e32 v1, v3, v2
+; GFX90A-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX90A-NEXT: .LBB0_5: ; %Flow1
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: .LBB0_6: ; %Flow2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execz .LBB0_8
+; GFX90A-NEXT: ; %bb.7: ; %atomicrmw.shared
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX90A-NEXT: ds_add_rtn_f32 v3, v0, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: .LBB0_8: ; %atomicrmw.phi
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_f32:
; GFX908: ; %bb.0:
; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT: flat_load_dword v3, v[0:1]
@@ -85,10 +128,10 @@ define <2 x half> @flat_agent_atomic_fadd_ret_v2f16(ptr %ptr, <2 x half> %val) {
; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX908-NEXT: v_mov_b32_e32 v4, v3
-; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX908-NEXT: v_add_f32_e32 v3, v4, v2
; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
+; GFX908-NEXT: buffer_wbinvl1
; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
@@ -98,1151 +141,1120 @@ define <2 x half> @flat_agent_atomic_fadd_ret_v2f16(ptr %ptr, <2 x half> %val) {
; GFX908-NEXT: v_mov_b32_e32 v0, v3
; GFX908-NEXT: s_setpc_b64 s[30:31]
;
-; GFX90A-LABEL: flat_agent_atomic_fadd_ret_v2f16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: flat_load_dword v3, v[0:1]
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB0_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB0_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB0_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
;
-; GFX940-LABEL: flat_agent_atomic_fadd_ret_v2f16:
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB0_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr %ptr, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @flat_agent_atomic_fadd_ret_f32__offset12b_pos(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_pos:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: flat_load_dword v3, v[0:1]
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB0_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 offset:2044 sc0
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB0_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX10-LABEL: flat_agent_atomic_fadd_ret_v2f16:
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_pos:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB1_1: ; %atomicrmw.start
; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v4, v3
-; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_add_f32_e32 v0, v1, v2
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX10-NEXT: buffer_gl1_inv
; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB0_1
+; GFX10-NEXT: s_cbranch_execnz .LBB1_1
; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: v_mov_b32_e32 v0, v3
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
-; GFX11-LABEL: flat_agent_atomic_fadd_ret_v2f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: flat_load_b32 v3, v[0:1]
-; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: .LBB0_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
-; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB0_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_mov_b32_e32 v0, v3
-; GFX11-NEXT: s_setpc_b64 s[30:31]
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fc, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v1, vcc
+; GFX90A-NEXT: s_mov_b64 s[4:5], src_shared_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s5, v5
+; GFX90A-NEXT: ; implicit-def: $vgpr0
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX90A-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB1_3
+; GFX90A-NEXT: ; %bb.1: ; %Flow2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB1_8
+; GFX90A-NEXT: .LBB1_2: ; %atomicrmw.phi
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+; GFX90A-NEXT: .LBB1_3: ; %atomicrmw.check.private
+; GFX90A-NEXT: s_mov_b64 s[6:7], src_private_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s7, v5
+; GFX90A-NEXT: ; implicit-def: $vgpr0
+; GFX90A-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX90A-NEXT: s_xor_b64 s[6:7], exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB1_5
+; GFX90A-NEXT: ; %bb.4: ; %atomicrmw.global
+; GFX90A-NEXT: global_atomic_add_f32 v0, v[4:5], v2, off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: ; implicit-def: $vgpr4_vgpr5
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: .LBB1_5: ; %Flow
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[6:7], s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB1_7
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.private
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v1, -1, v4, vcc
+; GFX90A-NEXT: buffer_load_dword v0, v1, s[0:3], 0 offen
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f32_e32 v2, v0, v2
+; GFX90A-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; GFX90A-NEXT: .LBB1_7: ; %Flow1
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: ; implicit-def: $vgpr4_vgpr5
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execz .LBB1_2
+; GFX90A-NEXT: .LBB1_8: ; %atomicrmw.shared
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, -1, v4, vcc
+; GFX90A-NEXT: ds_add_rtn_f32 v0, v0, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
-; GFX12-LABEL: flat_agent_atomic_fadd_ret_v2f16:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: flat_load_b32 v3, v[0:1]
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB0_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB0_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: v_mov_b32_e32 v0, v3
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr %ptr, <2 x half> %val syncscope("agent") seq_cst
- ret <2 x half> %result
-}
-
-define void @flat_agent_atomic_fadd_noret_v2f16(ptr %ptr, <2 x half> %val) {
-; GFX7-LABEL: flat_agent_atomic_fadd_noret_v2f16:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: flat_load_dword v5, v[0:1]
-; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
-; GFX7-NEXT: .LBB1_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
-; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
-; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
-; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
-; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
-; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
-; GFX7-NEXT: flat_atomic_cmpswap v7, v[0:1], v[5:6] glc
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: buffer_wbinvl1_vol
-; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
-; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
-; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB1_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: flat_agent_atomic_fadd_noret_v2f16:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: flat_load_dword v4, v[0:1]
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB1_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_add_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
-; GFX8-NEXT: v_add_f16_e32 v5, v4, v2
-; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
-; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: buffer_wbinvl1_vol
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v4, v3
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB1_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: flat_agent_atomic_fadd_noret_v2f16:
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_pos:
; GFX908: ; %bb.0:
; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
; GFX908-NEXT: s_mov_b64 s[4:5], 0
; GFX908-NEXT: .LBB1_1: ; %atomicrmw.start
; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
+; GFX908-NEXT: buffer_wbinvl1
; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v4, v3
; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT: s_cbranch_execnz .LBB1_1
; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
; GFX908-NEXT: s_setpc_b64 s[30:31]
;
-; GFX90A-LABEL: flat_agent_atomic_fadd_noret_v2f16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: flat_load_dword v5, v[0:1]
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB1_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB1_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB1_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
;
-; GFX940-LABEL: flat_agent_atomic_fadd_noret_v2f16:
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v1, v0
+; GFX7-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB1_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %result = atomicrmw fadd ptr %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @flat_agent_atomic_fadd_ret_f32__offset12b_neg(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_neg:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: flat_load_dword v5, v[0:1]
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB1_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 offset:63488 sc0
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB1_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX10-LABEL: flat_agent_atomic_fadd_noret_v2f16:
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 offset:63488 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_neg:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB2_1: ; %atomicrmw.start
; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_add_f32_e32 v0, v1, v2
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX10-NEXT: buffer_gl1_inv
; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB1_1
+; GFX10-NEXT: s_cbranch_execnz .LBB2_1
; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
-; GFX11-LABEL: flat_agent_atomic_fadd_noret_v2f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: flat_load_b32 v4, v[0:1]
-; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: .LBB1_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
-; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX11-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB1_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: s_mov_b64 s[4:5], src_shared_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s5, v5
+; GFX90A-NEXT: ; implicit-def: $vgpr0
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX90A-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB2_3
+; GFX90A-NEXT: ; %bb.1: ; %Flow2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB2_8
+; GFX90A-NEXT: .LBB2_2: ; %atomicrmw.phi
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+; GFX90A-NEXT: .LBB2_3: ; %atomicrmw.check.private
+; GFX90A-NEXT: s_mov_b64 s[6:7], src_private_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s7, v5
+; GFX90A-NEXT: ; implicit-def: $vgpr0
+; GFX90A-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX90A-NEXT: s_xor_b64 s[6:7], exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB2_5
+; GFX90A-NEXT: ; %bb.4: ; %atomicrmw.global
+; GFX90A-NEXT: global_atomic_add_f32 v0, v[4:5], v2, off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: ; implicit-def: $vgpr4_vgpr5
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: .LBB2_5: ; %Flow
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[6:7], s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB2_7
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.private
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v1, -1, v4, vcc
+; GFX90A-NEXT: buffer_load_dword v0, v1, s[0:3], 0 offen
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f32_e32 v2, v0, v2
+; GFX90A-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; GFX90A-NEXT: .LBB2_7: ; %Flow1
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: ; implicit-def: $vgpr4_vgpr5
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execz .LBB2_2
+; GFX90A-NEXT: .LBB2_8: ; %atomicrmw.shared
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, -1, v4, vcc
+; GFX90A-NEXT: ds_add_rtn_f32 v0, v0, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
-; GFX12-LABEL: flat_agent_atomic_fadd_noret_v2f16:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: flat_load_b32 v4, v[0:1]
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB1_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB1_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr %ptr, <2 x half> %val syncscope("agent") seq_cst
- ret void
-}
-
-define <2 x bfloat> @flat_agent_atomic_fadd_ret_v2bf16(ptr %ptr, <2 x bfloat> %val) {
-; GFX7-LABEL: flat_agent_atomic_fadd_ret_v2bf16:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: flat_load_dword v5, v[0:1]
-; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v3
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
-; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
-; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX7-NEXT: .LBB2_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
-; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
-; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GFX7-NEXT: v_add_f32_e32 v7, v7, v5
-; GFX7-NEXT: v_add_f32_e32 v6, v6, v4
-; GFX7-NEXT: v_alignbit_b32 v3, v2, v3, 16
-; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v7
-; GFX7-NEXT: v_alignbit_b32 v2, v2, v6, 16
-; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[2:3] glc
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: buffer_wbinvl1_vol
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
-; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB2_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: v_mov_b32_e32 v0, v3
-; GFX7-NEXT: v_mov_b32_e32 v1, v2
-; GFX7-NEXT: s_setpc_b64 s[30:31]
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v0, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB2_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
;
-; GFX8-LABEL: flat_agent_atomic_fadd_ret_v2bf16:
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_neg:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: flat_load_dword v3, v[0:1]
-; GFX8-NEXT: s_mov_b64 s[6:7], 0
-; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
; GFX8-NEXT: .LBB2_1: ; %atomicrmw.start
; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_mov_b32_e32 v6, v3
-; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX8-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX8-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX8-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v3
-; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
-; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
-; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
-; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX8-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
-; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GFX8-NEXT: v_alignbit_b32 v5, v5, v3, 16
-; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: buffer_wbinvl1_vol
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
-; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT: s_cbranch_execnz .LBB2_1
; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
-; GFX908-LABEL: flat_agent_atomic_fadd_ret_v2bf16:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: flat_load_dword v3, v[0:1]
-; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB2_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v6, v3
-; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX908-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
-; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
-; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
-; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB2_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: v_mov_b32_e32 v0, v3
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: flat_agent_atomic_fadd_ret_v2bf16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: flat_load_dword v3, v[0:1]
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB2_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v7, v3
-; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
-; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
-; GFX90A-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
-; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
-; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB2_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v1, v0
+; GFX7-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 -512
+ %result = atomicrmw fadd ptr %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define void @flat_agent_atomic_fadd_noret_f32(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_add_f32 v[0:1], v2
+; GFX12-NEXT: s_wait_storecnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
;
-; GFX940-LABEL: flat_agent_atomic_fadd_ret_v2bf16:
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_f32:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: flat_load_dword v3, v[0:1]
-; GFX940-NEXT: s_mov_b64 s[2:3], 0
-; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB2_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v7, v3
-; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
-; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
-; GFX940-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
-; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
-; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
-; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] sc0
+; GFX940-NEXT: flat_atomic_add_f32 v[0:1], v2
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
-; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB2_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX10-LABEL: flat_agent_atomic_fadd_ret_v2bf16:
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_add_f32 v[0:1], v2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_f32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: flat_load_dword v3, v[0:1]
-; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX10-NEXT: s_mov_b32 s5, 0
-; GFX10-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB3_1: ; %atomicrmw.start
; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v6, v3
-; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX10-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
-; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
-; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX10-NEXT: buffer_gl1_inv
; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_cbranch_execnz .LBB2_1
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB3_1
; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
-; GFX11-LABEL: flat_agent_atomic_fadd_ret_v2bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: flat_load_b32 v3, v[0:1]
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: s_mov_b32 s1, 0
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB2_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v6, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
-; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] glc
-; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_cbranch_execnz .LBB2_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: v_mov_b32_e32 v0, v3
-; GFX11-NEXT: s_setpc_b64 s[30:31]
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_f32:
+; GFX90A: ; %bb.0: ; %atomicrmw.check.shared
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: s_mov_b64 s[4:5], src_shared_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s5, v1
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX90A-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB3_3
+; GFX90A-NEXT: ; %bb.1: ; %Flow2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB3_8
+; GFX90A-NEXT: .LBB3_2: ; %atomicrmw.phi
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+; GFX90A-NEXT: .LBB3_3: ; %atomicrmw.check.private
+; GFX90A-NEXT: s_mov_b64 s[6:7], src_private_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s7, v1
+; GFX90A-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX90A-NEXT: s_xor_b64 s[6:7], exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB3_5
+; GFX90A-NEXT: ; %bb.4: ; %atomicrmw.global
+; GFX90A-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: .LBB3_5: ; %Flow
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[6:7], s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB3_7
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.private
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX90A-NEXT: buffer_load_dword v1, v0, s[0:3], 0 offen
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f32_e32 v1, v1, v2
+; GFX90A-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX90A-NEXT: .LBB3_7: ; %Flow1
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execz .LBB3_2
+; GFX90A-NEXT: .LBB3_8: ; %atomicrmw.shared
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX90A-NEXT: ds_add_f32 v0, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
-; GFX12-LABEL: flat_agent_atomic_fadd_ret_v2bf16:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: flat_load_b32 v3, v[0:1]
-; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB2_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v6, v3
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX12-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
-; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB2_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: v_mov_b32_e32 v0, v3
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr %ptr, <2 x bfloat> %val syncscope("agent") seq_cst
- ret <2 x bfloat> %result
-}
-
-define void @flat_agent_atomic_fadd_noret_v2bf16(ptr %ptr, <2 x bfloat> %val) {
-; GFX7-LABEL: flat_agent_atomic_fadd_noret_v2bf16:
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_f32:
+; GFX908: ; %bb.0: ; %atomicrmw.check.shared
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: s_mov_b64 s[4:5], src_shared_base
+; GFX908-NEXT: v_cmp_ne_u32_e32 vcc, s5, v1
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX908-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB3_3
+; GFX908-NEXT: ; %bb.1: ; %Flow2
+; GFX908-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB3_8
+; GFX908-NEXT: .LBB3_2: ; %atomicrmw.phi
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+; GFX908-NEXT: .LBB3_3: ; %atomicrmw.check.private
+; GFX908-NEXT: s_mov_b64 s[6:7], src_private_base
+; GFX908-NEXT: v_cmp_ne_u32_e32 vcc, s7, v1
+; GFX908-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX908-NEXT: s_xor_b64 s[6:7], exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execz .LBB3_5
+; GFX908-NEXT: ; %bb.4: ; %atomicrmw.global
+; GFX908-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX908-NEXT: ; implicit-def: $vgpr2
+; GFX908-NEXT: .LBB3_5: ; %Flow
+; GFX908-NEXT: s_andn2_saveexec_b64 s[6:7], s[6:7]
+; GFX908-NEXT: s_cbranch_execz .LBB3_7
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.private
+; GFX908-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX908-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX908-NEXT: buffer_load_dword v1, v0, s[0:3], 0 offen
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f32_e32 v1, v1, v2
+; GFX908-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX908-NEXT: .LBB3_7: ; %Flow1
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX908-NEXT: ; implicit-def: $vgpr2
+; GFX908-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_cbranch_execz .LBB3_2
+; GFX908-NEXT: .LBB3_8: ; %atomicrmw.shared
+; GFX908-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX908-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX908-NEXT: ds_add_f32 v0, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB3_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_f32:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: flat_load_dword v5, v[0:1]
-; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
-; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
; GFX7-NEXT: .LBB3_1: ; %atomicrmw.start
; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
-; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
-; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
-; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
-; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
-; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX7-NEXT: v_alignbit_b32 v5, v4, v5, 16
-; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v7
-; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
-; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[4:5] glc
; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: buffer_wbinvl1_vol
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
-; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT: s_cbranch_execnz .LBB3_1
; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fadd ptr %ptr, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fadd_noret_f32__offset12b_pos(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_add_f32 v[0:1], v2 offset:2044
+; GFX12-NEXT: s_wait_storecnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
;
-; GFX8-LABEL: flat_agent_atomic_fadd_noret_v2bf16:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: flat_load_dword v3, v[0:1]
-; GFX8-NEXT: s_mov_b64 s[6:7], 0
-; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX8-NEXT: .LBB3_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX8-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX8-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
-; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
-; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
-; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
-; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
-; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: buffer_wbinvl1_vol
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX8-NEXT: v_mov_b32_e32 v3, v2
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX8-NEXT: s_cbranch_execnz .LBB3_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX8-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90-LABEL: flat_agent_atomic_fadd_noret_v2bf16:
-; GFX90: ; %bb.0:
-; GFX90-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90-NEXT: flat_load_dword v3, v[0:1]
-; GFX90-NEXT: s_mov_b64 s[6:7], 0
-; GFX90-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX90-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX90-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90-NEXT: .LBB3_1: ; %atomicrmw.start
-; GFX90-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX90-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX90-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX90-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX90-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX90-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX90-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX90-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX90-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX90-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX90-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX90-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX90-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX90-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX90-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX90-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
-; GFX90-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90-NEXT: buffer_wbinvl1_vol
-; GFX90-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX90-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90-NEXT: v_mov_b32_e32 v3, v2
-; GFX90-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90-NEXT: s_cbranch_execnz .LBB3_1
-; GFX90-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: flat_agent_atomic_fadd_noret_v2bf16:
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_pos:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: flat_load_dword v3, v[0:1]
-; GFX940-NEXT: s_mov_b64 s[2:3], 0
-; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB3_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX940-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX940-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
-; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
-; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: flat_atomic_add_f32 v[0:1], v2 offset:2044
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB3_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX10-LABEL: flat_agent_atomic_fadd_noret_v2bf16:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: flat_load_dword v3, v[0:1]
-; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX10-NEXT: s_mov_b32 s5, 0
-; GFX10-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_add_f32 v[0:1], v2 offset:2044
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: .LBB4_1: ; %atomicrmw.start
; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX10-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX10-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
-; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
-; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX10-NEXT: buffer_gl1_inv
; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX10-NEXT: v_mov_b32_e32 v3, v2
-; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_cbranch_execnz .LBB3_1
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB4_1
; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
-; GFX11-LABEL: flat_agent_atomic_fadd_noret_v2bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: flat_load_b32 v3, v[0:1]
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX11-NEXT: s_mov_b32 s1, 0
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB3_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX11-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
-; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX11-NEXT: v_mov_b32_e32 v3, v2
-; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_cbranch_execnz .LBB3_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_setpc_b64 s[30:31]
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_pos:
+; GFX90A: ; %bb.0: ; %atomicrmw.check.shared
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0x7fc, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: s_mov_b64 s[4:5], src_shared_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s5, v1
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX90A-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB4_3
+; GFX90A-NEXT: ; %bb.1: ; %Flow2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB4_8
+; GFX90A-NEXT: .LBB4_2: ; %atomicrmw.phi
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+; GFX90A-NEXT: .LBB4_3: ; %atomicrmw.check.private
+; GFX90A-NEXT: s_mov_b64 s[6:7], src_private_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s7, v1
+; GFX90A-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX90A-NEXT: s_xor_b64 s[6:7], exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB4_5
+; GFX90A-NEXT: ; %bb.4: ; %atomicrmw.global
+; GFX90A-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: .LBB4_5: ; %Flow
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[6:7], s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB4_7
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.private
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX90A-NEXT: buffer_load_dword v1, v0, s[0:3], 0 offen
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f32_e32 v1, v1, v2
+; GFX90A-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX90A-NEXT: .LBB4_7: ; %Flow1
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execz .LBB4_2
+; GFX90A-NEXT: .LBB4_8: ; %atomicrmw.shared
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX90A-NEXT: ds_add_f32 v0, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
-; GFX12-LABEL: flat_agent_atomic_fadd_noret_v2bf16:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: flat_load_b32 v3, v[0:1]
-; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB3_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX12-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX12-NEXT: v_mov_b32_e32 v3, v2
-; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB3_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr %ptr, <2 x bfloat> %val syncscope("agent") seq_cst
- ret void
-}
-
-define <2 x half> @flat_system_atomic_fadd_ret_v2f16(ptr %ptr, <2 x half> %val) {
-; GFX7-LABEL: flat_system_atomic_fadd_ret_v2f16:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: flat_load_dword v5, v[0:1]
-; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v3
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v6
-; GFX7-NEXT: .LBB4_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
-; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX7-NEXT: v_add_f32_e32 v6, v6, v4
-; GFX7-NEXT: v_add_f32_e32 v7, v7, v5
-; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
-; GFX7-NEXT: v_or_b32_e32 v7, v2, v3
-; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v6
-; GFX7-NEXT: v_or_b32_e32 v6, v8, v2
-; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[6:7] glc
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: buffer_wbinvl1_vol
-; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v6
-; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v6
-; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v7
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB4_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: v_mov_b32_e32 v0, v2
-; GFX7-NEXT: v_mov_b32_e32 v1, v3
-; GFX7-NEXT: s_setpc_b64 s[30:31]
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_pos:
+; GFX908: ; %bb.0: ; %atomicrmw.check.shared
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0x7fc, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: s_mov_b64 s[4:5], src_shared_base
+; GFX908-NEXT: v_cmp_ne_u32_e32 vcc, s5, v1
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX908-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB4_3
+; GFX908-NEXT: ; %bb.1: ; %Flow2
+; GFX908-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB4_8
+; GFX908-NEXT: .LBB4_2: ; %atomicrmw.phi
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+; GFX908-NEXT: .LBB4_3: ; %atomicrmw.check.private
+; GFX908-NEXT: s_mov_b64 s[6:7], src_private_base
+; GFX908-NEXT: v_cmp_ne_u32_e32 vcc, s7, v1
+; GFX908-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX908-NEXT: s_xor_b64 s[6:7], exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execz .LBB4_5
+; GFX908-NEXT: ; %bb.4: ; %atomicrmw.global
+; GFX908-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX908-NEXT: ; implicit-def: $vgpr2
+; GFX908-NEXT: .LBB4_5: ; %Flow
+; GFX908-NEXT: s_andn2_saveexec_b64 s[6:7], s[6:7]
+; GFX908-NEXT: s_cbranch_execz .LBB4_7
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.private
+; GFX908-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX908-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX908-NEXT: buffer_load_dword v1, v0, s[0:3], 0 offen
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f32_e32 v1, v1, v2
+; GFX908-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX908-NEXT: .LBB4_7: ; %Flow1
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX908-NEXT: ; implicit-def: $vgpr2
+; GFX908-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_cbranch_execz .LBB4_2
+; GFX908-NEXT: .LBB4_8: ; %atomicrmw.shared
+; GFX908-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX908-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX908-NEXT: ds_add_f32 v0, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: s_setpc_b64 s[30:31]
;
-; GFX8-LABEL: flat_system_atomic_fadd_ret_v2f16:
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_pos:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
; GFX8-NEXT: s_mov_b64 s[4:5], 0
; GFX8-NEXT: .LBB4_1: ; %atomicrmw.start
; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_mov_b32_e32 v4, v3
-; GFX8-NEXT: v_add_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
-; GFX8-NEXT: v_add_f16_e32 v5, v4, v2
-; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_add_f32_e32 v3, v4, v2
; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: buffer_wbinvl1_vol
+; GFX8-NEXT: buffer_wbinvl1
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT: s_cbranch_execnz .LBB4_1
; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v0, v3
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
-; GFX908-LABEL: flat_system_atomic_fadd_ret_v2f16:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: flat_load_dword v3, v[0:1]
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB4_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v4, v3
-; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB4_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v0, v3
-; GFX908-NEXT: s_setpc_b64 s[30:31]
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB4_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %unused = atomicrmw fadd ptr %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fadd_noret_f32__offset12b_neg(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_add_f32 v[0:1], v2 offset:-2048
+; GFX12-NEXT: s_wait_storecnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
;
-; GFX90A-LABEL: flat_system_atomic_fadd_ret_v2f16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: flat_load_dword v3, v[0:1]
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB4_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX90A-NEXT: buffer_wbl2
-; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: buffer_invl2
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB4_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: flat_system_atomic_fadd_ret_v2f16:
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_neg:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: flat_load_dword v3, v[0:1]
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB4_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX940-NEXT: buffer_wbl2 sc0 sc1
-; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0 sc1
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_add_f32 v[0:1], v2 offset:63488
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: buffer_inv sc0 sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB4_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: buffer_inv sc1
; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX10-LABEL: flat_system_atomic_fadd_ret_v2f16:
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_add_f32 v[0:1], v2 offset:63488
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_neg:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: .LBB5_1: ; %atomicrmw.start
; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v4, v3
-; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX10-NEXT: buffer_gl1_inv
; GFX10-NEXT: buffer_gl0_inv
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB4_1
+; GFX10-NEXT: s_cbranch_execnz .LBB5_1
; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: v_mov_b32_e32 v0, v3
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
-; GFX11-LABEL: flat_system_atomic_fadd_ret_v2f16:
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_neg:
+; GFX90A: ; %bb.0: ; %atomicrmw.check.shared
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: s_mov_b64 s[4:5], src_shared_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s5, v1
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX90A-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB5_3
+; GFX90A-NEXT: ; %bb.1: ; %Flow2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB5_8
+; GFX90A-NEXT: .LBB5_2: ; %atomicrmw.phi
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+; GFX90A-NEXT: .LBB5_3: ; %atomicrmw.check.private
+; GFX90A-NEXT: s_mov_b64 s[6:7], src_private_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s7, v1
+; GFX90A-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX90A-NEXT: s_xor_b64 s[6:7], exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB5_5
+; GFX90A-NEXT: ; %bb.4: ; %atomicrmw.global
+; GFX90A-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: .LBB5_5: ; %Flow
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[6:7], s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB5_7
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.private
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX90A-NEXT: buffer_load_dword v1, v0, s[0:3], 0 offen
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f32_e32 v1, v1, v2
+; GFX90A-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX90A-NEXT: .LBB5_7: ; %Flow1
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execz .LBB5_2
+; GFX90A-NEXT: .LBB5_8: ; %atomicrmw.shared
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX90A-NEXT: ds_add_f32 v0, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_neg:
+; GFX908: ; %bb.0: ; %atomicrmw.check.shared
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: s_mov_b64 s[4:5], src_shared_base
+; GFX908-NEXT: v_cmp_ne_u32_e32 vcc, s5, v1
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX908-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_3
+; GFX908-NEXT: ; %bb.1: ; %Flow2
+; GFX908-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_8
+; GFX908-NEXT: .LBB5_2: ; %atomicrmw.phi
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+; GFX908-NEXT: .LBB5_3: ; %atomicrmw.check.private
+; GFX908-NEXT: s_mov_b64 s[6:7], src_private_base
+; GFX908-NEXT: v_cmp_ne_u32_e32 vcc, s7, v1
+; GFX908-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX908-NEXT: s_xor_b64 s[6:7], exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execz .LBB5_5
+; GFX908-NEXT: ; %bb.4: ; %atomicrmw.global
+; GFX908-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX908-NEXT: ; implicit-def: $vgpr2
+; GFX908-NEXT: .LBB5_5: ; %Flow
+; GFX908-NEXT: s_andn2_saveexec_b64 s[6:7], s[6:7]
+; GFX908-NEXT: s_cbranch_execz .LBB5_7
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.private
+; GFX908-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX908-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX908-NEXT: buffer_load_dword v1, v0, s[0:3], 0 offen
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f32_e32 v1, v1, v2
+; GFX908-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX908-NEXT: .LBB5_7: ; %Flow1
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX908-NEXT: ; implicit-def: $vgpr2
+; GFX908-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_cbranch_execz .LBB5_2
+; GFX908-NEXT: .LBB5_8: ; %atomicrmw.shared
+; GFX908-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX908-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX908-NEXT: ds_add_f32 v0, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 -512
+ %unused = atomicrmw fadd ptr %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define float @flat_system_atomic_fadd_ret_f32__offset12b_pos(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fadd_ret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB6_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fadd_ret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fadd_ret_f32__offset12b_pos:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX11-NEXT: .LBB6_1: ; %atomicrmw.start
; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_mov_b32_e32 v4, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX11-NEXT: v_add_f32_e32 v3, v4, v2
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT: buffer_gl1_inv
; GFX11-NEXT: buffer_gl0_inv
@@ -1250,1427 +1262,1880 @@ define <2 x half> @flat_system_atomic_fadd_ret_v2f16(ptr %ptr, <2 x half> %val)
; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB4_1
+; GFX11-NEXT: s_cbranch_execnz .LBB6_1
; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT: v_mov_b32_e32 v0, v3
; GFX11-NEXT: s_setpc_b64 s[30:31]
;
-; GFX12-LABEL: flat_system_atomic_fadd_ret_v2f16:
+; GFX10-LABEL: flat_system_atomic_fadd_ret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB6_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fadd_ret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB6_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fadd_ret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB6_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fadd_ret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB6_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fadd_ret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v1, v0
+; GFX7-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB6_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %result = atomicrmw fadd ptr %gep, float %val seq_cst
+ ret float %result
+}
+
+define void @flat_system_atomic_fadd_noret_f32__offset12b_pos(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fadd_noret_f32__offset12b_pos:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT: s_wait_expcnt 0x0
; GFX12-NEXT: s_wait_samplecnt 0x0
; GFX12-NEXT: s_wait_bvhcnt 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: flat_load_b32 v4, v[0:1] offset:2044
; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX12-NEXT: .LBB7_1: ; %atomicrmw.start
; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX12-NEXT: v_add_f32_e32 v3, v4, v2
; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT: global_inv scope:SCOPE_SYS
; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB4_1
+; GFX12-NEXT: s_cbranch_execnz .LBB7_1
; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: v_mov_b32_e32 v0, v3
; GFX12-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr %ptr, <2 x half> %val seq_cst
- ret <2 x half> %result
-}
-
-define void @flat_system_atomic_fadd_noret_v2f16(ptr %ptr, <2 x half> %val) {
-; GFX7-LABEL: flat_system_atomic_fadd_noret_v2f16:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: flat_load_dword v5, v[0:1]
-; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
-; GFX7-NEXT: .LBB5_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
-; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
-; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
-; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
-; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
-; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
-; GFX7-NEXT: flat_atomic_cmpswap v7, v[0:1], v[5:6] glc
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: buffer_wbinvl1_vol
-; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
-; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
-; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB5_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_setpc_b64 s[30:31]
;
-; GFX8-LABEL: flat_system_atomic_fadd_noret_v2f16:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: flat_load_dword v4, v[0:1]
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB5_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_add_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
-; GFX8-NEXT: v_add_f16_e32 v5, v4, v2
-; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
-; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: buffer_wbinvl1_vol
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v4, v3
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB5_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_setpc_b64 s[30:31]
+; GFX940-LABEL: flat_system_atomic_fadd_noret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_add_f32 v[0:1], v2 offset:2044 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX908-LABEL: flat_system_atomic_fadd_noret_v2f16:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: flat_load_dword v4, v[0:1]
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB5_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v4, v3
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB5_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
+; GFX11-LABEL: flat_system_atomic_fadd_noret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v4, v[0:1] offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB7_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fadd_noret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB7_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
;
-; GFX90A-LABEL: flat_system_atomic_fadd_noret_v2f16:
+; GFX90A-LABEL: flat_system_atomic_fadd_noret_f32__offset12b_pos:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: flat_load_dword v5, v[0:1] offset:2044
; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX90A-NEXT: .LBB7_1: ; %atomicrmw.start
; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX90A-NEXT: v_add_f32_e32 v4, v5, v2
; GFX90A-NEXT: buffer_wbl2
-; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] glc
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX90A-NEXT: buffer_invl2
-; GFX90A-NEXT: buffer_wbinvl1_vol
+; GFX90A-NEXT: buffer_wbinvl1
; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT: v_mov_b32_e32 v5, v3
; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB5_1
+; GFX90A-NEXT: s_cbranch_execnz .LBB7_1
; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
-; GFX940-LABEL: flat_system_atomic_fadd_noret_v2f16:
+; GFX908-LABEL: flat_system_atomic_fadd_noret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v4, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB7_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fadd_noret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB7_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fadd_noret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB7_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %unused = atomicrmw fadd ptr %gep, float %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; float with ftz/daz
+; --------------------------------------------------------------------
+
+define float @flat_agent_atomic_fadd_ret_f32__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_f32__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_f32__ftz:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: flat_load_dword v5, v[0:1]
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB5_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
-; GFX940-NEXT: buffer_wbl2 sc0 sc1
-; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0 sc1
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 sc0
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: buffer_inv sc0 sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB5_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: buffer_inv sc1
; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX10-LABEL: flat_system_atomic_fadd_noret_v2f16:
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_f32__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_f32__ftz:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX10-NEXT: .LBB8_1: ; %atomicrmw.start
; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX10-NEXT: buffer_gl1_inv
; GFX10-NEXT: buffer_gl0_inv
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX10-NEXT: v_mov_b32_e32 v4, v3
; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB5_1
+; GFX10-NEXT: s_cbranch_execnz .LBB8_1
; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
-; GFX11-LABEL: flat_system_atomic_fadd_noret_v2f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: flat_load_b32 v4, v[0:1]
-; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: .LBB5_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
-; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX11-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB5_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_f32__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: s_mov_b64 s[4:5], src_shared_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s5, v1
+; GFX90A-NEXT: ; implicit-def: $vgpr3
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX90A-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execz .LBB8_6
+; GFX90A-NEXT: ; %bb.1: ; %atomicrmw.check.private
+; GFX90A-NEXT: s_mov_b64 s[6:7], src_private_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s7, v1
+; GFX90A-NEXT: ; implicit-def: $vgpr3
+; GFX90A-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX90A-NEXT: s_xor_b64 s[6:7], exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB8_3
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.global
+; GFX90A-NEXT: global_atomic_add_f32 v3, v[0:1], v2, off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: .LBB8_3: ; %Flow
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[6:7], s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB8_5
+; GFX90A-NEXT: ; %bb.4: ; %atomicrmw.private
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX90A-NEXT: buffer_load_dword v3, v0, s[0:3], 0 offen
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f32_e32 v1, v3, v2
+; GFX90A-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX90A-NEXT: .LBB8_5: ; %Flow1
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: .LBB8_6: ; %Flow2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execz .LBB8_8
+; GFX90A-NEXT: ; %bb.7: ; %atomicrmw.shared
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX90A-NEXT: ds_add_rtn_f32 v3, v0, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: .LBB8_8: ; %atomicrmw.phi
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
-; GFX12-LABEL: flat_system_atomic_fadd_noret_v2f16:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: flat_load_b32 v4, v[0:1]
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB5_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SYS
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB5_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr %ptr, <2 x half> %val seq_cst
- ret void
-}
-
-define <2 x bfloat> @flat_system_atomic_fadd_ret_v2bf16(ptr %ptr, <2 x bfloat> %val) {
-; GFX7-LABEL: flat_system_atomic_fadd_ret_v2bf16:
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_f32__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_f32__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_f32__ftz:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: flat_load_dword v5, v[0:1]
-; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v3
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
-; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
-; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX7-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX7-NEXT: .LBB8_1: ; %atomicrmw.start
; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
-; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
-; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
-; GFX7-NEXT: v_add_f32_e32 v7, v7, v5
-; GFX7-NEXT: v_add_f32_e32 v6, v6, v4
-; GFX7-NEXT: v_alignbit_b32 v3, v2, v3, 16
-; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v7
-; GFX7-NEXT: v_alignbit_b32 v2, v2, v6, 16
-; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[2:3] glc
; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: buffer_wbinvl1_vol
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
-; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v6
; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB6_1
+; GFX7-NEXT: s_cbranch_execnz .LBB8_1
; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT: v_mov_b32_e32 v0, v3
-; GFX7-NEXT: v_mov_b32_e32 v1, v2
; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr %ptr, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @flat_agent_atomic_fadd_ret_f32__offset12b_pos__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
;
-; GFX8-LABEL: flat_system_atomic_fadd_ret_v2bf16:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: flat_load_dword v3, v[0:1]
-; GFX8-NEXT: s_mov_b64 s[6:7], 0
-; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX8-NEXT: .LBB6_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_mov_b32_e32 v6, v3
-; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX8-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX8-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX8-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v3
-; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
-; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
-; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
-; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX8-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
-; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GFX8-NEXT: v_alignbit_b32 v5, v5, v3, 16
-; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: buffer_wbinvl1_vol
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
-; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX8-NEXT: s_cbranch_execnz .LBB6_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX8-NEXT: v_mov_b32_e32 v0, v3
-; GFX8-NEXT: s_setpc_b64 s[30:31]
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX908-LABEL: flat_system_atomic_fadd_ret_v2bf16:
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB9_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fc, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v1, vcc
+; GFX90A-NEXT: s_mov_b64 s[4:5], src_shared_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s5, v5
+; GFX90A-NEXT: ; implicit-def: $vgpr0
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX90A-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB9_3
+; GFX90A-NEXT: ; %bb.1: ; %Flow2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB9_8
+; GFX90A-NEXT: .LBB9_2: ; %atomicrmw.phi
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+; GFX90A-NEXT: .LBB9_3: ; %atomicrmw.check.private
+; GFX90A-NEXT: s_mov_b64 s[6:7], src_private_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s7, v5
+; GFX90A-NEXT: ; implicit-def: $vgpr0
+; GFX90A-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX90A-NEXT: s_xor_b64 s[6:7], exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB9_5
+; GFX90A-NEXT: ; %bb.4: ; %atomicrmw.global
+; GFX90A-NEXT: global_atomic_add_f32 v0, v[4:5], v2, off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: ; implicit-def: $vgpr4_vgpr5
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: .LBB9_5: ; %Flow
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[6:7], s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB9_7
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.private
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v1, -1, v4, vcc
+; GFX90A-NEXT: buffer_load_dword v0, v1, s[0:3], 0 offen
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f32_e32 v2, v0, v2
+; GFX90A-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; GFX90A-NEXT: .LBB9_7: ; %Flow1
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: ; implicit-def: $vgpr4_vgpr5
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execz .LBB9_2
+; GFX90A-NEXT: .LBB9_8: ; %atomicrmw.shared
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, -1, v4, vcc
+; GFX90A-NEXT: ds_add_rtn_f32 v0, v0, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_pos__ftz:
; GFX908: ; %bb.0:
; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: flat_load_dword v3, v[0:1]
-; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB9_1: ; %atomicrmw.start
; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v6, v3
-; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX908-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
-; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
-; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
-; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB6_1
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB9_1
; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT: v_mov_b32_e32 v0, v3
; GFX908-NEXT: s_setpc_b64 s[30:31]
;
-; GFX90A-LABEL: flat_system_atomic_fadd_ret_v2bf16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: flat_load_dword v3, v[0:1]
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB6_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v7, v3
-; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
-; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
-; GFX90A-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
-; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
-; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
-; GFX90A-NEXT: buffer_wbl2
-; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: buffer_invl2
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB6_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: flat_system_atomic_fadd_ret_v2bf16:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: flat_load_dword v3, v[0:1]
-; GFX940-NEXT: s_mov_b64 s[2:3], 0
-; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB6_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v7, v3
-; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
-; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
-; GFX940-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
-; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
-; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
-; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
-; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
-; GFX940-NEXT: buffer_wbl2 sc0 sc1
-; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] sc0 sc1
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB9_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v1, v0
+; GFX7-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB9_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %result = atomicrmw fadd ptr %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @flat_agent_atomic_fadd_ret_f32__offset12b_neg__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_neg__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_neg__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 offset:63488 sc0
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: buffer_inv sc0 sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
-; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB6_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: buffer_inv sc1
; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX10-LABEL: flat_system_atomic_fadd_ret_v2bf16:
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_neg__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 offset:63488 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_neg__ftz:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: flat_load_dword v3, v[0:1]
-; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX10-NEXT: s_mov_b32 s5, 0
-; GFX10-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB10_1: ; %atomicrmw.start
; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v6, v3
-; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX10-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
-; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
-; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_add_f32_e32 v0, v1, v2
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX10-NEXT: buffer_gl1_inv
; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_cbranch_execnz .LBB6_1
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB10_1
; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
-; GFX11-LABEL: flat_system_atomic_fadd_ret_v2bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: flat_load_b32 v3, v[0:1]
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT: s_mov_b32 s1, 0
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB6_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v6, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
-; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] glc
-; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_cbranch_execnz .LBB6_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: v_mov_b32_e32 v0, v3
-; GFX11-NEXT: s_setpc_b64 s[30:31]
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_neg__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: s_mov_b64 s[4:5], src_shared_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s5, v5
+; GFX90A-NEXT: ; implicit-def: $vgpr0
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX90A-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB10_3
+; GFX90A-NEXT: ; %bb.1: ; %Flow2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB10_8
+; GFX90A-NEXT: .LBB10_2: ; %atomicrmw.phi
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+; GFX90A-NEXT: .LBB10_3: ; %atomicrmw.check.private
+; GFX90A-NEXT: s_mov_b64 s[6:7], src_private_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s7, v5
+; GFX90A-NEXT: ; implicit-def: $vgpr0
+; GFX90A-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX90A-NEXT: s_xor_b64 s[6:7], exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB10_5
+; GFX90A-NEXT: ; %bb.4: ; %atomicrmw.global
+; GFX90A-NEXT: global_atomic_add_f32 v0, v[4:5], v2, off glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: ; implicit-def: $vgpr4_vgpr5
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: .LBB10_5: ; %Flow
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[6:7], s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB10_7
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.private
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v1, -1, v4, vcc
+; GFX90A-NEXT: buffer_load_dword v0, v1, s[0:3], 0 offen
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f32_e32 v2, v0, v2
+; GFX90A-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
+; GFX90A-NEXT: .LBB10_7: ; %Flow1
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: ; implicit-def: $vgpr4_vgpr5
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execz .LBB10_2
+; GFX90A-NEXT: .LBB10_8: ; %atomicrmw.shared
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, -1, v4, vcc
+; GFX90A-NEXT: ds_add_rtn_f32 v0, v0, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_neg__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v0, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB10_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_neg__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB10_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
;
-; GFX12-LABEL: flat_system_atomic_fadd_ret_v2bf16:
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_f32__offset12b_neg__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v1, v0
+; GFX7-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB10_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 -512
+ %result = atomicrmw fadd ptr %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define void @flat_agent_atomic_fadd_noret_f32__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_f32__ftz:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT: s_wait_expcnt 0x0
; GFX12-NEXT: s_wait_samplecnt 0x0
; GFX12-NEXT: s_wait_bvhcnt 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: flat_load_b32 v3, v[0:1]
-; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB6_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v6, v3
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX12-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
-; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SYS
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
-; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB6_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: flat_atomic_add_f32 v[0:1], v2
+; GFX12-NEXT: s_wait_storecnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
; GFX12-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr %ptr, <2 x bfloat> %val seq_cst
- ret <2 x bfloat> %result
-}
-
-define void @flat_system_atomic_fadd_noret_v2bf16(ptr %ptr, <2 x bfloat> %val) {
-; GFX7-LABEL: flat_system_atomic_fadd_noret_v2bf16:
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_f32__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_add_f32 v[0:1], v2
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_f32__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_add_f32 v[0:1], v2
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_f32__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB11_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_f32__ftz:
+; GFX90A: ; %bb.0: ; %atomicrmw.check.shared
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: s_mov_b64 s[4:5], src_shared_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s5, v1
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX90A-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_3
+; GFX90A-NEXT: ; %bb.1: ; %Flow2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_8
+; GFX90A-NEXT: .LBB11_2: ; %atomicrmw.phi
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+; GFX90A-NEXT: .LBB11_3: ; %atomicrmw.check.private
+; GFX90A-NEXT: s_mov_b64 s[6:7], src_private_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s7, v1
+; GFX90A-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX90A-NEXT: s_xor_b64 s[6:7], exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB11_5
+; GFX90A-NEXT: ; %bb.4: ; %atomicrmw.global
+; GFX90A-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: .LBB11_5: ; %Flow
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[6:7], s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB11_7
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.private
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX90A-NEXT: buffer_load_dword v1, v0, s[0:3], 0 offen
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f32_e32 v1, v1, v2
+; GFX90A-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX90A-NEXT: .LBB11_7: ; %Flow1
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execz .LBB11_2
+; GFX90A-NEXT: .LBB11_8: ; %atomicrmw.shared
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX90A-NEXT: ds_add_f32 v0, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_f32__ftz:
+; GFX908: ; %bb.0: ; %atomicrmw.check.shared
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: s_mov_b64 s[4:5], src_shared_base
+; GFX908-NEXT: v_cmp_ne_u32_e32 vcc, s5, v1
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX908-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_3
+; GFX908-NEXT: ; %bb.1: ; %Flow2
+; GFX908-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_8
+; GFX908-NEXT: .LBB11_2: ; %atomicrmw.phi
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+; GFX908-NEXT: .LBB11_3: ; %atomicrmw.check.private
+; GFX908-NEXT: s_mov_b64 s[6:7], src_private_base
+; GFX908-NEXT: v_cmp_ne_u32_e32 vcc, s7, v1
+; GFX908-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX908-NEXT: s_xor_b64 s[6:7], exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execz .LBB11_5
+; GFX908-NEXT: ; %bb.4: ; %atomicrmw.global
+; GFX908-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX908-NEXT: ; implicit-def: $vgpr2
+; GFX908-NEXT: .LBB11_5: ; %Flow
+; GFX908-NEXT: s_andn2_saveexec_b64 s[6:7], s[6:7]
+; GFX908-NEXT: s_cbranch_execz .LBB11_7
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.private
+; GFX908-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX908-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX908-NEXT: buffer_load_dword v1, v0, s[0:3], 0 offen
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f32_e32 v1, v1, v2
+; GFX908-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX908-NEXT: .LBB11_7: ; %Flow1
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX908-NEXT: ; implicit-def: $vgpr2
+; GFX908-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_cbranch_execz .LBB11_2
+; GFX908-NEXT: .LBB11_8: ; %atomicrmw.shared
+; GFX908-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX908-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX908-NEXT: ds_add_f32 v0, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_f32__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_f32__ftz:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: flat_load_dword v5, v[0:1]
-; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
-; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX7-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX7-NEXT: .LBB11_1: ; %atomicrmw.start
; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
-; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
-; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
-; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
-; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
-; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX7-NEXT: v_alignbit_b32 v5, v4, v5, 16
-; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v7
-; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
-; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[4:5] glc
; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: buffer_wbinvl1_vol
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
-; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB7_1
+; GFX7-NEXT: s_cbranch_execnz .LBB11_1
; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fadd ptr %ptr, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fadd_noret_f32__offset12b_pos__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_add_f32 v[0:1], v2 offset:2044
+; GFX12-NEXT: s_wait_storecnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_add_f32 v[0:1], v2 offset:2044
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_add_f32 v[0:1], v2 offset:2044
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB12_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0: ; %atomicrmw.check.shared
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0x7fc, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: s_mov_b64 s[4:5], src_shared_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s5, v1
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX90A-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB12_3
+; GFX90A-NEXT: ; %bb.1: ; %Flow2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB12_8
+; GFX90A-NEXT: .LBB12_2: ; %atomicrmw.phi
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+; GFX90A-NEXT: .LBB12_3: ; %atomicrmw.check.private
+; GFX90A-NEXT: s_mov_b64 s[6:7], src_private_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s7, v1
+; GFX90A-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX90A-NEXT: s_xor_b64 s[6:7], exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB12_5
+; GFX90A-NEXT: ; %bb.4: ; %atomicrmw.global
+; GFX90A-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: .LBB12_5: ; %Flow
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[6:7], s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB12_7
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.private
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX90A-NEXT: buffer_load_dword v1, v0, s[0:3], 0 offen
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f32_e32 v1, v1, v2
+; GFX90A-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX90A-NEXT: .LBB12_7: ; %Flow1
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execz .LBB12_2
+; GFX90A-NEXT: .LBB12_8: ; %atomicrmw.shared
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX90A-NEXT: ds_add_f32 v0, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0: ; %atomicrmw.check.shared
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0x7fc, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: s_mov_b64 s[4:5], src_shared_base
+; GFX908-NEXT: v_cmp_ne_u32_e32 vcc, s5, v1
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX908-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB12_3
+; GFX908-NEXT: ; %bb.1: ; %Flow2
+; GFX908-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB12_8
+; GFX908-NEXT: .LBB12_2: ; %atomicrmw.phi
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+; GFX908-NEXT: .LBB12_3: ; %atomicrmw.check.private
+; GFX908-NEXT: s_mov_b64 s[6:7], src_private_base
+; GFX908-NEXT: v_cmp_ne_u32_e32 vcc, s7, v1
+; GFX908-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX908-NEXT: s_xor_b64 s[6:7], exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execz .LBB12_5
+; GFX908-NEXT: ; %bb.4: ; %atomicrmw.global
+; GFX908-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX908-NEXT: ; implicit-def: $vgpr2
+; GFX908-NEXT: .LBB12_5: ; %Flow
+; GFX908-NEXT: s_andn2_saveexec_b64 s[6:7], s[6:7]
+; GFX908-NEXT: s_cbranch_execz .LBB12_7
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.private
+; GFX908-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX908-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX908-NEXT: buffer_load_dword v1, v0, s[0:3], 0 offen
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f32_e32 v1, v1, v2
+; GFX908-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX908-NEXT: .LBB12_7: ; %Flow1
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX908-NEXT: ; implicit-def: $vgpr2
+; GFX908-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_cbranch_execz .LBB12_2
+; GFX908-NEXT: .LBB12_8: ; %atomicrmw.shared
+; GFX908-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX908-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX908-NEXT: ds_add_f32 v0, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: s_setpc_b64 s[30:31]
;
-; GFX8-LABEL: flat_system_atomic_fadd_noret_v2bf16:
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_pos__ftz:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: flat_load_dword v3, v[0:1]
-; GFX8-NEXT: s_mov_b64 s[6:7], 0
-; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX8-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB12_1: ; %atomicrmw.start
; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX8-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX8-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
-; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
-; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
-; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
-; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
-; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: buffer_wbinvl1_vol
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX8-NEXT: v_mov_b32_e32 v3, v2
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX8-NEXT: s_cbranch_execnz .LBB7_1
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB12_1
; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
-; GFX908-LABEL: flat_system_atomic_fadd_noret_v2bf16:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: flat_load_dword v3, v[0:1]
-; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB7_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX908-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX908-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX908-NEXT: v_mov_b32_e32 v3, v2
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB7_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: flat_system_atomic_fadd_noret_v2bf16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: flat_load_dword v3, v[0:1]
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB7_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX90A-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX90A-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
-; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
-; GFX90A-NEXT: buffer_wbl2
-; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: buffer_invl2
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: v_mov_b32_e32 v3, v2
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB7_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB12_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %unused = atomicrmw fadd ptr %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fadd_noret_f32__offset12b_neg__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_neg__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_add_f32 v[0:1], v2 offset:-2048
+; GFX12-NEXT: s_wait_storecnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
;
-; GFX940-LABEL: flat_system_atomic_fadd_noret_v2bf16:
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_neg__ftz:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: flat_load_dword v3, v[0:1]
-; GFX940-NEXT: s_mov_b64 s[2:3], 0
-; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB7_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX940-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX940-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
-; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
-; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
-; GFX940-NEXT: buffer_wbl2 sc0 sc1
-; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0 sc1
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_add_f32 v[0:1], v2 offset:63488
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: buffer_inv sc0 sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB7_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: buffer_inv sc1
; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX10-LABEL: flat_system_atomic_fadd_noret_v2bf16:
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_neg__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_add_f32 v[0:1], v2 offset:63488
+; GFX11-NEXT: s_waitcnt lgkmcnt(0)
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_neg__ftz:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: flat_load_dword v3, v[0:1]
-; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX10-NEXT: s_mov_b32 s5, 0
-; GFX10-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: .LBB13_1: ; %atomicrmw.start
; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX10-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX10-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
-; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
-; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX10-NEXT: buffer_gl1_inv
; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX10-NEXT: v_mov_b32_e32 v3, v2
-; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_cbranch_execnz .LBB7_1
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB13_1
; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
-; GFX11-LABEL: flat_system_atomic_fadd_noret_v2bf16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: flat_load_b32 v3, v[0:1]
-; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX11-NEXT: s_mov_b32 s1, 0
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB7_1: ; %atomicrmw.start
-; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX11-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
-; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: buffer_gl1_inv
-; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX11-NEXT: v_mov_b32_e32 v3, v2
-; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_cbranch_execnz .LBB7_1
-; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_setpc_b64 s[30:31]
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_neg__ftz:
+; GFX90A: ; %bb.0: ; %atomicrmw.check.shared
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: s_mov_b64 s[4:5], src_shared_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s5, v1
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX90A-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB13_3
+; GFX90A-NEXT: ; %bb.1: ; %Flow2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB13_8
+; GFX90A-NEXT: .LBB13_2: ; %atomicrmw.phi
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+; GFX90A-NEXT: .LBB13_3: ; %atomicrmw.check.private
+; GFX90A-NEXT: s_mov_b64 s[6:7], src_private_base
+; GFX90A-NEXT: v_cmp_ne_u32_e32 vcc, s7, v1
+; GFX90A-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX90A-NEXT: s_xor_b64 s[6:7], exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB13_5
+; GFX90A-NEXT: ; %bb.4: ; %atomicrmw.global
+; GFX90A-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: .LBB13_5: ; %Flow
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[6:7], s[6:7]
+; GFX90A-NEXT: s_cbranch_execz .LBB13_7
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.private
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX90A-NEXT: buffer_load_dword v1, v0, s[0:3], 0 offen
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f32_e32 v1, v1, v2
+; GFX90A-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX90A-NEXT: .LBB13_7: ; %Flow1
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX90A-NEXT: ; implicit-def: $vgpr2
+; GFX90A-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_cbranch_execz .LBB13_2
+; GFX90A-NEXT: .LBB13_8: ; %atomicrmw.shared
+; GFX90A-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX90A-NEXT: ds_add_f32 v0, v2
+; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_neg__ftz:
+; GFX908: ; %bb.0: ; %atomicrmw.check.shared
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: s_mov_b64 s[4:5], src_shared_base
+; GFX908-NEXT: v_cmp_ne_u32_e32 vcc, s5, v1
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], vcc
+; GFX908-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB13_3
+; GFX908-NEXT: ; %bb.1: ; %Flow2
+; GFX908-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB13_8
+; GFX908-NEXT: .LBB13_2: ; %atomicrmw.phi
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+; GFX908-NEXT: .LBB13_3: ; %atomicrmw.check.private
+; GFX908-NEXT: s_mov_b64 s[6:7], src_private_base
+; GFX908-NEXT: v_cmp_ne_u32_e32 vcc, s7, v1
+; GFX908-NEXT: s_and_saveexec_b64 s[6:7], vcc
+; GFX908-NEXT: s_xor_b64 s[6:7], exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execz .LBB13_5
+; GFX908-NEXT: ; %bb.4: ; %atomicrmw.global
+; GFX908-NEXT: global_atomic_add_f32 v[0:1], v2, off
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX908-NEXT: ; implicit-def: $vgpr2
+; GFX908-NEXT: .LBB13_5: ; %Flow
+; GFX908-NEXT: s_andn2_saveexec_b64 s[6:7], s[6:7]
+; GFX908-NEXT: s_cbranch_execz .LBB13_7
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.private
+; GFX908-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX908-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX908-NEXT: buffer_load_dword v1, v0, s[0:3], 0 offen
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f32_e32 v1, v1, v2
+; GFX908-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX908-NEXT: .LBB13_7: ; %Flow1
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: ; implicit-def: $vgpr0_vgpr1
+; GFX908-NEXT: ; implicit-def: $vgpr2
+; GFX908-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_cbranch_execz .LBB13_2
+; GFX908-NEXT: .LBB13_8: ; %atomicrmw.shared
+; GFX908-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[0:1]
+; GFX908-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc
+; GFX908-NEXT: ds_add_f32 v0, v2
+; GFX908-NEXT: s_waitcnt lgkmcnt(0)
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_neg__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB13_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
;
-; GFX12-LABEL: flat_system_atomic_fadd_noret_v2bf16:
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_f32__offset12b_neg__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB13_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 -512
+ %unused = atomicrmw fadd ptr %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define float @flat_system_atomic_fadd_ret_f32__offset12b_pos__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_system_atomic_fadd_ret_f32__offset12b_pos__ftz:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT: s_wait_expcnt 0x0
; GFX12-NEXT: s_wait_samplecnt 0x0
; GFX12-NEXT: s_wait_bvhcnt 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: flat_load_b32 v3, v[0:1]
-; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB14_1: ; %atomicrmw.start
; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX12-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: v_add_f32_e32 v3, v4, v2
; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT: global_inv scope:SCOPE_SYS
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX12-NEXT: v_mov_b32_e32 v3, v2
-; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB7_1
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB14_1
; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
; GFX12-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr %ptr, <2 x bfloat> %val seq_cst
- ret void
-}
-
-define <4 x half> @flat_agent_atomic_fadd_ret_v4f16(ptr %ptr, <4 x half> %val) {
-; GFX7-LABEL: flat_agent_atomic_fadd_ret_v4f16:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_add_i32_e32 v6, vcc, 4, v0
-; GFX7-NEXT: flat_load_dword v9, v[0:1]
-; GFX7-NEXT: v_addc_u32_e32 v7, vcc, 0, v1, vcc
-; GFX7-NEXT: flat_load_dword v10, v[6:7]
-; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GFX7-NEXT: v_cvt_f16_f32_e32 v11, v4
-; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v8, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v9
-; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v9
-; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v10
-; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v10
-; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v9, v11
-; GFX7-NEXT: .LBB8_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GFX7-NEXT: v_cvt_f16_f32_e32 v10, v4
-; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v11, v5
-; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GFX7-NEXT: v_or_b32_e32 v4, v10, v4
-; GFX7-NEXT: v_or_b32_e32 v5, v2, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GFX7-NEXT: v_add_f32_e32 v11, v11, v6
-; GFX7-NEXT: v_add_f32_e32 v3, v3, v8
-; GFX7-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GFX7-NEXT: v_add_f32_e32 v10, v10, v7
-; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GFX7-NEXT: v_add_f32_e32 v2, v2, v9
-; GFX7-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GFX7-NEXT: v_cvt_f16_f32_e32 v12, v2
-; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v11
-; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX7-NEXT: v_or_b32_e32 v2, v10, v2
-; GFX7-NEXT: v_or_b32_e32 v3, v12, v3
-; GFX7-NEXT: flat_atomic_cmpswap_x2 v[10:11], v[0:1], v[2:5] glc
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: buffer_wbinvl1_vol
-; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[10:11], v[4:5]
-; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v10
-; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v11
-; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v11
-; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v10
-; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB8_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: v_mov_b32_e32 v0, v4
-; GFX7-NEXT: v_mov_b32_e32 v1, v5
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: flat_agent_atomic_fadd_ret_v4f16:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_add_u32_e32 v5, vcc, 4, v0
-; GFX8-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
-; GFX8-NEXT: flat_load_dword v4, v[0:1]
-; GFX8-NEXT: flat_load_dword v5, v[5:6]
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB8_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_mov_b32_e32 v7, v5
-; GFX8-NEXT: v_mov_b32_e32 v6, v4
-; GFX8-NEXT: v_add_f16_sdwa v4, v7, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
-; GFX8-NEXT: v_add_f16_e32 v5, v7, v3
-; GFX8-NEXT: v_add_f16_sdwa v8, v6, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
-; GFX8-NEXT: v_add_f16_e32 v9, v6, v2
-; GFX8-NEXT: v_or_b32_e32 v5, v5, v4
-; GFX8-NEXT: v_or_b32_e32 v4, v9, v8
-; GFX8-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: buffer_wbinvl1_vol
-; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB8_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v0, v4
-; GFX8-NEXT: v_mov_b32_e32 v1, v5
-; GFX8-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: flat_agent_atomic_fadd_ret_v4f16:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB8_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v7, v5
-; GFX908-NEXT: v_mov_b32_e32 v6, v4
-; GFX908-NEXT: v_pk_add_f16 v5, v7, v3
-; GFX908-NEXT: v_pk_add_f16 v4, v6, v2
-; GFX908-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB8_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v0, v4
-; GFX908-NEXT: v_mov_b32_e32 v1, v5
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: flat_agent_atomic_fadd_ret_v4f16:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB8_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
-; GFX90A-NEXT: v_pk_add_f16 v5, v7, v3
-; GFX90A-NEXT: v_pk_add_f16 v4, v6, v2
-; GFX90A-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB8_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v4
-; GFX90A-NEXT: v_mov_b32_e32 v1, v5
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
-; GFX940-LABEL: flat_agent_atomic_fadd_ret_v4f16:
+; GFX940-LABEL: flat_system_atomic_fadd_ret_f32__offset12b_pos__ftz:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB8_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
-; GFX940-NEXT: v_pk_add_f16 v5, v7, v3
-; GFX940-NEXT: v_pk_add_f16 v4, v6, v2
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] sc0
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_add_f32 v0, v[0:1], v2 offset:2044 sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB8_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v0, v4
-; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: buffer_inv sc0 sc1
; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX10-LABEL: flat_agent_atomic_fadd_ret_v4f16:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
-; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: .LBB8_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v7, v5
-; GFX10-NEXT: v_mov_b32_e32 v6, v4
-; GFX10-NEXT: v_pk_add_f16 v5, v7, v3
-; GFX10-NEXT: v_pk_add_f16 v4, v6, v2
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
-; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB8_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: v_mov_b32_e32 v0, v4
-; GFX10-NEXT: v_mov_b32_e32 v1, v5
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: flat_agent_atomic_fadd_ret_v4f16:
+; GFX11-LABEL: flat_system_atomic_fadd_ret_f32__offset12b_pos__ftz:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: flat_load_b64 v[4:5], v[0:1]
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX11-NEXT: .LBB14_1: ; %atomicrmw.start
; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_pk_add_f16 v5, v7, v3
-; GFX11-NEXT: v_pk_add_f16 v4, v6, v2
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v3, v4, v2
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] glc
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT: buffer_gl1_inv
; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB8_1
+; GFX11-NEXT: s_cbranch_execnz .LBB14_1
; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
; GFX11-NEXT: s_setpc_b64 s[30:31]
;
-; GFX12-LABEL: flat_agent_atomic_fadd_ret_v4f16:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: flat_load_b64 v[4:5], v[0:1]
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB8_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-NEXT: v_pk_add_f16 v5, v7, v3
-; GFX12-NEXT: v_pk_add_f16 v4, v6, v2
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB8_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr %ptr, <4 x half> %val syncscope("agent") seq_cst
- ret <4 x half> %result
-}
-
-define <4 x half> @flat_system_atomic_fadd_ret_v4f16(ptr %ptr, <4 x half> %val) {
-; GFX7-LABEL: flat_system_atomic_fadd_ret_v4f16:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_add_i32_e32 v6, vcc, 4, v0
-; GFX7-NEXT: flat_load_dword v9, v[0:1]
-; GFX7-NEXT: v_addc_u32_e32 v7, vcc, 0, v1, vcc
-; GFX7-NEXT: flat_load_dword v10, v[6:7]
-; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GFX7-NEXT: v_cvt_f16_f32_e32 v11, v4
-; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v8, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v9
-; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v9
-; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v10
-; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v10
-; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v9, v11
-; GFX7-NEXT: .LBB9_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GFX7-NEXT: v_cvt_f16_f32_e32 v10, v4
-; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
-; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v11, v5
-; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GFX7-NEXT: v_or_b32_e32 v4, v10, v4
-; GFX7-NEXT: v_or_b32_e32 v5, v2, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v10, v10
-; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
-; GFX7-NEXT: v_add_f32_e32 v11, v11, v6
-; GFX7-NEXT: v_add_f32_e32 v3, v3, v8
-; GFX7-NEXT: v_cvt_f16_f32_e32 v11, v11
-; GFX7-NEXT: v_add_f32_e32 v10, v10, v7
-; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GFX7-NEXT: v_add_f32_e32 v2, v2, v9
-; GFX7-NEXT: v_cvt_f16_f32_e32 v10, v10
-; GFX7-NEXT: v_cvt_f16_f32_e32 v12, v2
-; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v11
-; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX7-NEXT: v_or_b32_e32 v2, v10, v2
-; GFX7-NEXT: v_or_b32_e32 v3, v12, v3
-; GFX7-NEXT: flat_atomic_cmpswap_x2 v[10:11], v[0:1], v[2:5] glc
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: buffer_wbinvl1_vol
-; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[10:11], v[4:5]
-; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v10
-; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v11
-; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v11
-; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v10
-; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB9_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: v_mov_b32_e32 v0, v4
-; GFX7-NEXT: v_mov_b32_e32 v1, v5
-; GFX7-NEXT: s_setpc_b64 s[30:31]
+; GFX10-LABEL: flat_system_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB14_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
;
-; GFX8-LABEL: flat_system_atomic_fadd_ret_v4f16:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_add_u32_e32 v5, vcc, 4, v0
-; GFX8-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
-; GFX8-NEXT: flat_load_dword v4, v[0:1]
-; GFX8-NEXT: flat_load_dword v5, v[5:6]
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB9_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_mov_b32_e32 v7, v5
-; GFX8-NEXT: v_mov_b32_e32 v6, v4
-; GFX8-NEXT: v_add_f16_sdwa v4, v7, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
-; GFX8-NEXT: v_add_f16_e32 v5, v7, v3
-; GFX8-NEXT: v_add_f16_sdwa v8, v6, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
-; GFX8-NEXT: v_add_f16_e32 v9, v6, v2
-; GFX8-NEXT: v_or_b32_e32 v5, v5, v4
-; GFX8-NEXT: v_or_b32_e32 v4, v9, v8
-; GFX8-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: buffer_wbinvl1_vol
-; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB9_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v0, v4
-; GFX8-NEXT: v_mov_b32_e32 v1, v5
-; GFX8-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: flat_system_atomic_fadd_ret_v4f16:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB9_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v7, v5
-; GFX908-NEXT: v_mov_b32_e32 v6, v4
-; GFX908-NEXT: v_pk_add_f16 v5, v7, v3
-; GFX908-NEXT: v_pk_add_f16 v4, v6, v2
-; GFX908-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB9_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v0, v4
-; GFX908-NEXT: v_mov_b32_e32 v1, v5
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: flat_system_atomic_fadd_ret_v4f16:
+; GFX90A-LABEL: flat_system_atomic_fadd_ret_f32__offset12b_pos__ftz:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX90A-NEXT: .LBB14_1: ; %atomicrmw.start
; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
-; GFX90A-NEXT: v_pk_add_f16 v5, v7, v3
-; GFX90A-NEXT: v_pk_add_f16 v4, v6, v2
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_add_f32_e32 v4, v5, v2
; GFX90A-NEXT: buffer_wbl2
-; GFX90A-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX90A-NEXT: buffer_invl2
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB9_1
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_1
; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v0, v4
-; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
-; GFX940-LABEL: flat_system_atomic_fadd_ret_v4f16:
+; GFX908-LABEL: flat_system_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fadd_ret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v1, v0
+; GFX7-NEXT: v_add_f32_e32 v0, v1, v2
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %result = atomicrmw fadd ptr %gep, float %val seq_cst
+ ret float %result
+}
+
+define void @flat_system_atomic_fadd_noret_f32__offset12b_pos__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_system_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v4, v[0:1] offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB15_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fadd_noret_f32__offset12b_pos__ftz:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB9_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
-; GFX940-NEXT: v_pk_add_f16 v5, v7, v3
-; GFX940-NEXT: v_pk_add_f16 v4, v6, v2
; GFX940-NEXT: buffer_wbl2 sc0 sc1
-; GFX940-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] sc0 sc1
+; GFX940-NEXT: flat_atomic_add_f32 v[0:1], v2 offset:2044 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX940-NEXT: buffer_inv sc0 sc1
-; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB9_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v0, v4
-; GFX940-NEXT: v_mov_b32_e32 v1, v5
; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX10-LABEL: flat_system_atomic_fadd_ret_v4f16:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
-; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: .LBB9_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v7, v5
-; GFX10-NEXT: v_mov_b32_e32 v6, v4
-; GFX10-NEXT: v_pk_add_f16 v5, v7, v3
-; GFX10-NEXT: v_pk_add_f16 v4, v6, v2
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
-; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB9_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: v_mov_b32_e32 v0, v4
-; GFX10-NEXT: v_mov_b32_e32 v1, v5
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: flat_system_atomic_fadd_ret_v4f16:
+; GFX11-LABEL: flat_system_atomic_fadd_noret_f32__offset12b_pos__ftz:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: flat_load_b64 v[4:5], v[0:1]
+; GFX11-NEXT: flat_load_b32 v4, v[0:1] offset:2044
; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX11-NEXT: .LBB15_1: ; %atomicrmw.start
; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_pk_add_f16 v5, v7, v3
-; GFX11-NEXT: v_pk_add_f16 v4, v6, v2
+; GFX11-NEXT: v_add_f32_e32 v3, v4, v2
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] glc
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT: buffer_gl1_inv
; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB9_1
+; GFX11-NEXT: s_cbranch_execnz .LBB15_1
; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
; GFX11-NEXT: s_setpc_b64 s[30:31]
;
-; GFX12-LABEL: flat_system_atomic_fadd_ret_v4f16:
+; GFX10-LABEL: flat_system_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB15_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v5, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB15_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v4, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB15_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB15_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fadd_noret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v3, v4, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB15_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %unused = atomicrmw fadd ptr %gep, float %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; double
+; --------------------------------------------------------------------
+
+define double @flat_agent_atomic_fadd_ret_f64(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_f64:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT: s_wait_expcnt 0x0
@@ -2679,971 +3144,9565 @@ define <4 x half> @flat_system_atomic_fadd_ret_v4f16(ptr %ptr, <4 x half> %val)
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: flat_load_b64 v[4:5], v[0:1]
; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX12-NEXT: .LBB16_1: ; %atomicrmw.start
; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-NEXT: v_pk_add_f16 v5, v7, v3
-; GFX12-NEXT: v_pk_add_f16 v4, v6, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_add_f64_e32 v[4:5], v[6:7], v[2:3]
; GFX12-NEXT: s_wait_storecnt 0x0
; GFX12-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] th:TH_ATOMIC_RETURN
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB9_1
+; GFX12-NEXT: s_cbranch_execnz .LBB16_1
; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
; GFX12-NEXT: s_setpc_b64 s[30:31]
- %result = atomicrmw fadd ptr %ptr, <4 x half> %val seq_cst
- ret <4 x half> %result
-}
-
-define <2 x half> @flat_agent_atomic_fadd_ret_v2f16_offset(ptr %ptr, <2 x half> %val) {
-; GFX7-LABEL: flat_agent_atomic_fadd_ret_v2f16_offset:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x3ffc, v0
-; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
-; GFX7-NEXT: flat_load_dword v1, v[4:5]
-; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v3
-; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v2
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v0
-; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
-; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GFX7-NEXT: .LBB10_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v1
-; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v0
-; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
-; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
-; GFX7-NEXT: v_or_b32_e32 v7, v0, v1
-; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
-; GFX7-NEXT: v_or_b32_e32 v6, v8, v0
-; GFX7-NEXT: flat_atomic_cmpswap v6, v[4:5], v[6:7] glc
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: buffer_wbinvl1_vol
-; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v6
-; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v6
-; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v7
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB10_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: flat_agent_atomic_fadd_ret_v2f16_offset:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x3ffc, v0
-; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
-; GFX8-NEXT: flat_load_dword v0, v[3:4]
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB10_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_mov_b32_e32 v1, v0
-; GFX8-NEXT: v_add_f16_sdwa v0, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
-; GFX8-NEXT: v_add_f16_e32 v5, v1, v2
-; GFX8-NEXT: v_or_b32_e32 v0, v5, v0
-; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: buffer_wbinvl1_vol
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB10_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: flat_agent_atomic_fadd_ret_v2f16_offset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x3ffc, v0
-; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, 0, v1, vcc
-; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0x3000, v0
-; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX908-NEXT: flat_load_dword v0, v[0:1] offset:4092
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB10_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v1, v0
-; GFX908-NEXT: v_pk_add_f16 v0, v1, v2
-; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB10_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: flat_agent_atomic_fadd_ret_v2f16_offset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x3ffc, v0
-; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v1, vcc
-; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0x3000, v0
-; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX90A-NEXT: flat_load_dword v0, v[0:1] offset:4092
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB10_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v1, v0
-; GFX90A-NEXT: v_pk_add_f16 v0, v1, v2
-; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB10_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
-; GFX940-LABEL: flat_agent_atomic_fadd_ret_v2f16_offset:
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_f64:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v4, v0
-; GFX940-NEXT: v_mov_b32_e32 v5, v1
-; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0x3000, v4
-; GFX940-NEXT: s_mov_b64 s[0:1], 0x3ffc
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v5, vcc
-; GFX940-NEXT: flat_load_dword v0, v[0:1] offset:4092
-; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB10_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v1, v0
-; GFX940-NEXT: v_pk_add_f16 v0, v1, v2
; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] sc0
+; GFX940-NEXT: flat_atomic_add_f64 v[0:1], v[0:1], v[2:3] sc0
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB10_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX10-LABEL: flat_agent_atomic_fadd_ret_v2f16_offset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x3ffc, v0
-; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
-; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: flat_load_dword v0, v[3:4]
-; GFX10-NEXT: .LBB10_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v1, v0
-; GFX10-NEXT: v_pk_add_f16 v0, v1, v2
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
-; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB10_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: flat_agent_atomic_fadd_ret_v2f16_offset:
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_f64:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: flat_load_b64 v[4:5], v[0:1]
; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x3000, v3
-; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, 0, v1, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x3ffc, v3
-; GFX11-NEXT: flat_load_b32 v0, v[4:5] offset:4092
-; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
-; GFX11-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX11-NEXT: .LBB16_1: ; %atomicrmw.start
; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v1, v0
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_pk_add_f16 v0, v1, v2
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: flat_atomic_cmpswap_b32 v0, v[3:4], v[0:1] glc
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] glc
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT: buffer_gl1_inv
; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB10_1
+; GFX11-NEXT: s_cbranch_execnz .LBB16_1
; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
; GFX11-NEXT: s_setpc_b64 s[30:31]
;
-; GFX12-LABEL: flat_agent_atomic_fadd_ret_v2f16_offset:
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB16_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_atomic_add_f64 v[0:1], v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB16_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 4, v0
+; GFX8-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: flat_load_dword v5, v[5:6]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v7, v5
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB16_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: v_mov_b32_e32 v1, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v5, vcc, 4, v0
+; GFX7-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: flat_load_dword v5, v[5:6]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB16_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v4
+; GFX7-NEXT: v_mov_b32_e32 v1, v5
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr %ptr, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define double @flat_agent_atomic_fadd_ret_f64__offset12b_pos(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_f64__offset12b_pos:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT: s_wait_expcnt 0x0
; GFX12-NEXT: s_wait_samplecnt 0x0
; GFX12-NEXT: s_wait_bvhcnt 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:16380
+; GFX12-NEXT: flat_load_b64 v[4:5], v[0:1] offset:2040
; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX12-NEXT: .LBB17_1: ; %atomicrmw.start
; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX12-NEXT: v_add_f64_e32 v[4:5], v[6:7], v[2:3]
; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:16380 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] offset:2040 th:TH_ATOMIC_RETURN
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB10_1
+; GFX12-NEXT: s_cbranch_execnz .LBB17_1
; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x half>, ptr %ptr, i64 4095
- %result = atomicrmw fadd ptr %gep, <2 x half> %val syncscope("agent") seq_cst
- ret <2 x half> %result
-}
-
-define <2 x half> @flat_agent_atomic_fadd_ret_v2f16_negoffset(ptr %ptr, <2 x half> %val) {
-; GFX7-LABEL: flat_agent_atomic_fadd_ret_v2f16_negoffset:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
-; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
-; GFX7-NEXT: flat_load_dword v1, v[4:5]
-; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v3
-; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v2
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v0
-; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
-; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GFX7-NEXT: .LBB11_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
-; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
-; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v1
-; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v0
-; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
-; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
-; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
-; GFX7-NEXT: v_or_b32_e32 v7, v0, v1
-; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
-; GFX7-NEXT: v_or_b32_e32 v6, v8, v0
-; GFX7-NEXT: flat_atomic_cmpswap v6, v[4:5], v[6:7] glc
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: buffer_wbinvl1_vol
-; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v6
-; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v6
-; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v7
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB11_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: flat_agent_atomic_fadd_ret_v2f16_negoffset:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
-; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
-; GFX8-NEXT: flat_load_dword v0, v[3:4]
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB11_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_mov_b32_e32 v1, v0
-; GFX8-NEXT: v_add_f16_sdwa v0, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
-; GFX8-NEXT: v_add_f16_e32 v5, v1, v2
-; GFX8-NEXT: v_or_b32_e32 v0, v5, v0
-; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: buffer_wbinvl1_vol
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB11_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: flat_agent_atomic_fadd_ret_v2f16_negoffset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
-; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
-; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
-; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
-; GFX908-NEXT: flat_load_dword v0, v[0:1]
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB11_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v1, v0
-; GFX908-NEXT: v_pk_add_f16 v0, v1, v2
-; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB11_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: flat_agent_atomic_fadd_ret_v2f16_negoffset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
-; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
-; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
-; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
-; GFX90A-NEXT: flat_load_dword v0, v[0:1]
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB11_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v1, v0
-; GFX90A-NEXT: v_pk_add_f16 v0, v1, v2
-; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB11_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
-; GFX940-LABEL: flat_agent_atomic_fadd_ret_v2f16_negoffset:
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_f64__offset12b_pos:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v4, v0
-; GFX940-NEXT: v_mov_b32_e32 v5, v1
-; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v4
-; GFX940-NEXT: s_movk_i32 s0, 0xf800
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v5, vcc
-; GFX940-NEXT: flat_load_dword v0, v[0:1]
-; GFX940-NEXT: s_mov_b32 s1, -1
-; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB11_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v1, v0
-; GFX940-NEXT: v_pk_add_f16 v0, v1, v2
; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] sc0
+; GFX940-NEXT: flat_atomic_add_f64 v[0:1], v[0:1], v[2:3] offset:2040 sc0
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB11_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX10-LABEL: flat_agent_atomic_fadd_ret_v2f16_negoffset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
-; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
-; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: flat_load_dword v0, v[3:4]
-; GFX10-NEXT: .LBB11_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v1, v0
-; GFX10-NEXT: v_pk_add_f16 v0, v1, v2
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
-; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB11_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: flat_agent_atomic_fadd_ret_v2f16_negoffset:
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_f64__offset12b_pos:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: flat_load_b64 v[4:5], v[0:1] offset:2040
; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v3
-; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v3
-; GFX11-NEXT: flat_load_b32 v0, v[4:5]
-; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
-; GFX11-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX11-NEXT: .LBB17_1: ; %atomicrmw.start
; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v1, v0
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_pk_add_f16 v0, v1, v2
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: flat_atomic_cmpswap_b32 v0, v[3:4], v[0:1] glc
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] offset:2040 glc
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT: buffer_gl1_inv
; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB11_1
+; GFX11-NEXT: s_cbranch_execnz .LBB17_1
; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
; GFX11-NEXT: s_setpc_b64 s[30:31]
;
-; GFX12-LABEL: flat_agent_atomic_fadd_ret_v2f16_negoffset:
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_f64__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7f8, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dwordx2 v[0:1], v[4:5]
+; GFX10-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v9, v1
+; GFX10-NEXT: v_mov_b32_e32 v8, v0
+; GFX10-NEXT: v_add_f64 v[6:7], v[8:9], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[8:9]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB17_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_f64__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_atomic_add_f64 v[0:1], v[0:1], v[2:3] offset:2040 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_f64__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dwordx2 v[4:5], v[0:1] offset:2040
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] offset:2040 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_f64__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7f8, v0
+; GFX8-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v1, v[0:1]
+; GFX8-NEXT: flat_load_dword v0, v[4:5]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v9, v1
+; GFX8-NEXT: v_mov_b32_e32 v8, v0
+; GFX8-NEXT: v_add_f64 v[6:7], v[8:9], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_f64__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7f8, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v1, v[0:1]
+; GFX7-NEXT: flat_load_dword v0, v[4:5]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v9, v1
+; GFX7-NEXT: v_mov_b32_e32 v8, v0
+; GFX7-NEXT: v_add_f64 v[6:7], v[8:9], v[2:3]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr %ptr, i64 255
+ %result = atomicrmw fadd ptr %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define double @flat_agent_atomic_fadd_ret_f64__offset12b_neg(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_f64__offset12b_neg:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT: s_wait_expcnt 0x0
; GFX12-NEXT: s_wait_samplecnt 0x0
; GFX12-NEXT: s_wait_bvhcnt 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: flat_load_b64 v[4:5], v[0:1] offset:-2048
; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX12-NEXT: .LBB18_1: ; %atomicrmw.start
; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX12-NEXT: v_add_f64_e32 v[4:5], v[6:7], v[2:3]
; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] offset:-2048 th:TH_ATOMIC_RETURN
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB11_1
+; GFX12-NEXT: s_cbranch_execnz .LBB18_1
; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x half>, ptr %ptr, i64 -512
- %result = atomicrmw fadd ptr %gep, <2 x half> %val syncscope("agent") seq_cst
- ret <2 x half> %result
-}
-
-define void @flat_agent_atomic_fadd_noret_v2f16_offset(ptr %ptr, <2 x half> %val) {
-; GFX7-LABEL: flat_agent_atomic_fadd_noret_v2f16_offset:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x3ffc, v0
-; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GFX7-NEXT: flat_load_dword v5, v[0:1]
-; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
-; GFX7-NEXT: .LBB12_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
-; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
-; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
-; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
-; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
-; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
-; GFX7-NEXT: flat_atomic_cmpswap v7, v[0:1], v[5:6] glc
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: buffer_wbinvl1_vol
-; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
-; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
-; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB12_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: flat_agent_atomic_fadd_noret_v2f16_offset:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x3ffc, v0
-; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GFX8-NEXT: flat_load_dword v4, v[0:1]
-; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB12_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_add_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
-; GFX8-NEXT: v_add_f16_e32 v5, v4, v2
-; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
-; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: buffer_wbinvl1_vol
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v4, v3
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB12_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: flat_agent_atomic_fadd_noret_v2f16_offset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x3ffc, v0
-; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, 0, v1, vcc
-; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0x3000, v0
-; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX908-NEXT: flat_load_dword v1, v[0:1] offset:4092
-; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB12_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_pk_add_f16 v0, v1, v2
-; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v1, v0
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB12_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: flat_agent_atomic_fadd_noret_v2f16_offset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x3ffc, v0
-; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v1, vcc
-; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0x3000, v0
-; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX90A-NEXT: flat_load_dword v1, v[0:1] offset:4092
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB12_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_pk_add_f16 v0, v1, v2
-; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v1, v0
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB12_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
-; GFX940-LABEL: flat_agent_atomic_fadd_noret_v2f16_offset:
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_f64__offset12b_neg:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0x3000, v0
-; GFX940-NEXT: s_mov_b64 s[0:1], 0x3ffc
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v1, vcc
-; GFX940-NEXT: flat_load_dword v5, v[4:5] offset:4092
-; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB12_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: flat_atomic_add_f64 v[0:1], v[0:1], v[2:3] offset:63488 sc0
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB12_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX10-LABEL: flat_agent_atomic_fadd_noret_v2f16_offset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x3ffc, v0
-; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
-; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: flat_load_dword v4, v[0:1]
-; GFX10-NEXT: .LBB12_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX10-NEXT: v_mov_b32_e32 v4, v3
-; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB12_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: flat_agent_atomic_fadd_noret_v2f16_offset:
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_f64__offset12b_neg:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x3000, v0
-; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0x3ffc, v0
-; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
-; GFX11-NEXT: flat_load_b32 v4, v[3:4] offset:4092
+; GFX11-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v5, v0
; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v5
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v4, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v5, vcc_lo, 0xfffff800, v5
+; GFX11-NEXT: v_add_co_ci_u32_e32 v6, vcc_lo, -1, v4, vcc_lo
+; GFX11-NEXT: flat_load_b64 v[0:1], v[0:1]
+; GFX11-NEXT: .LBB18_1: ; %atomicrmw.start
; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX11-NEXT: v_dual_mov_b32 v10, v1 :: v_dual_mov_b32 v9, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_f64 v[7:8], v[9:10], v[2:3]
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[0:1], v[5:6], v[7:10] glc
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT: buffer_gl1_inv
; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[9:10]
; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB12_1
+; GFX11-NEXT: s_cbranch_execnz .LBB18_1
; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT: s_setpc_b64 s[30:31]
;
-; GFX12-LABEL: flat_agent_atomic_fadd_noret_v2f16_offset:
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_f64__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dwordx2 v[0:1], v[4:5]
+; GFX10-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v9, v1
+; GFX10-NEXT: v_mov_b32_e32 v8, v0
+; GFX10-NEXT: v_add_f64 v[6:7], v[8:9], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[8:9]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB18_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_f64__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_atomic_add_f64 v[0:1], v[0:1], v[2:3] offset:63488 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_f64__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dwordx2 v[0:1], v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v9, v1
+; GFX908-NEXT: v_mov_b32_e32 v8, v0
+; GFX908-NEXT: v_add_f64 v[6:7], v[8:9], v[2:3]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB18_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_f64__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff804, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v1, v[0:1]
+; GFX8-NEXT: flat_load_dword v0, v[4:5]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v9, v1
+; GFX8-NEXT: v_mov_b32_e32 v8, v0
+; GFX8-NEXT: v_add_f64 v[6:7], v[8:9], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB18_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_f64__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff804, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v1, v[0:1]
+; GFX7-NEXT: flat_load_dword v0, v[4:5]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v9, v1
+; GFX7-NEXT: v_mov_b32_e32 v8, v0
+; GFX7-NEXT: v_add_f64 v[6:7], v[8:9], v[2:3]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB18_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr %ptr, i64 -256
+ %result = atomicrmw fadd ptr %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define void @flat_agent_atomic_fadd_noret_f64(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_f64:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT: s_wait_expcnt 0x0
; GFX12-NEXT: s_wait_samplecnt 0x0
; GFX12-NEXT: s_wait_bvhcnt 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: flat_load_b32 v4, v[0:1] offset:16380
+; GFX12-NEXT: flat_load_b64 v[6:7], v[0:1]
; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX12-NEXT: .LBB19_1: ; %atomicrmw.start
; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX12-NEXT: v_add_f64_e32 v[4:5], v[6:7], v[2:3]
; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:16380 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] th:TH_ATOMIC_RETURN
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB12_1
+; GFX12-NEXT: s_cbranch_execnz .LBB19_1
; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x half>, ptr %ptr, i64 4095
- %result = atomicrmw fadd ptr %gep, <2 x half> %val syncscope("agent") seq_cst
- ret void
-}
-
-define void @flat_agent_atomic_fadd_noret_v2f16_negoffset(ptr %ptr, <2 x half> %val) {
-; GFX7-LABEL: flat_agent_atomic_fadd_noret_v2f16_negoffset:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
-; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
-; GFX7-NEXT: flat_load_dword v5, v[0:1]
-; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
-; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
-; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
-; GFX7-NEXT: .LBB13_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
-; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
-; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
-; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
-; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
-; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
-; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
-; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
-; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
-; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
-; GFX7-NEXT: flat_atomic_cmpswap v7, v[0:1], v[5:6] glc
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: buffer_wbinvl1_vol
-; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
-; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
-; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB13_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_setpc_b64 s[30:31]
;
-; GFX8-LABEL: flat_agent_atomic_fadd_noret_v2f16_negoffset:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
-; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
-; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_add_f64 v[0:1], v[2:3]
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b64 v[6:7], v[0:1]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB19_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dwordx2 v[6:7], v[0:1]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB19_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_atomic_add_f64 v[0:1], v[2:3]
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dwordx2 v[6:7], v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB19_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 4, v0
+; GFX8-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v6, v[0:1]
+; GFX8-NEXT: flat_load_dword v7, v[4:5]
; GFX8-NEXT: s_mov_b64 s[4:5], 0
-; GFX8-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX8-NEXT: .LBB19_1: ; %atomicrmw.start
; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_add_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
-; GFX8-NEXT: v_add_f16_e32 v5, v4, v2
-; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
-; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: buffer_wbinvl1_vol
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v7, v5
; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX8-NEXT: s_cbranch_execnz .LBB13_1
+; GFX8-NEXT: s_cbranch_execnz .LBB19_1
; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
-; GFX908-LABEL: flat_agent_atomic_fadd_noret_v2f16_negoffset:
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 4, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v6, v[0:1]
+; GFX7-NEXT: flat_load_dword v7, v[4:5]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB19_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fadd ptr %ptr, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fadd_noret_f64__offset12b_pos(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_f64__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b64 v[6:7], v[0:1] offset:2040
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_add_f64_e32 v[4:5], v[6:7], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] offset:2040 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB20_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_f64__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_add_f64 v[0:1], v[2:3] offset:2040
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_f64__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b64 v[6:7], v[0:1] offset:2040
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] offset:2040 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB20_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_f64__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7f8, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dwordx2 v[6:7], v[0:1]
+; GFX10-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB20_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_f64__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_atomic_add_f64 v[0:1], v[2:3] offset:2040
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_f64__offset12b_pos:
; GFX908: ; %bb.0:
; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
-; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
-; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
-; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
-; GFX908-NEXT: flat_load_dword v1, v[0:1]
+; GFX908-NEXT: flat_load_dwordx2 v[6:7], v[0:1] offset:2040
; GFX908-NEXT: s_mov_b64 s[4:5], 0
-; GFX908-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX908-NEXT: .LBB20_1: ; %atomicrmw.start
; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_pk_add_f16 v0, v1, v2
-; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] offset:2040 glc
; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX908-NEXT: s_cbranch_execnz .LBB13_1
+; GFX908-NEXT: s_cbranch_execnz .LBB20_1
; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT: s_setpc_b64 s[30:31]
;
-; GFX90A-LABEL: flat_agent_atomic_fadd_noret_v2f16_negoffset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
-; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
-; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
-; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
-; GFX90A-NEXT: flat_load_dword v1, v[0:1]
-; GFX90A-NEXT: s_mov_b64 s[4:5], 0
-; GFX90A-NEXT: .LBB13_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_pk_add_f16 v0, v1, v2
-; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX90A-NEXT: v_mov_b32_e32 v1, v0
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_cbranch_execnz .LBB13_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_f64__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7f8, v0
+; GFX8-NEXT: v_addc_u32_e32 v9, vcc, 0, v1, vcc
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v7, v[0:1]
+; GFX8-NEXT: flat_load_dword v6, v[8:9]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[8:9], v[4:7] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v7, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB20_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_f64__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v8, vcc, 0x7f8, v0
+; GFX7-NEXT: v_addc_u32_e32 v9, vcc, 0, v1, vcc
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v7, v[0:1]
+; GFX7-NEXT: flat_load_dword v6, v[8:9]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[8:9], v[4:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v7, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB20_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr %ptr, i64 255
+ %unused = atomicrmw fadd ptr %gep, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fadd_noret_f64__offset12b_neg(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_f64__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b64 v[6:7], v[0:1] offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_add_f64_e32 v[4:5], v[6:7], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB21_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_f64__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_add_f64 v[0:1], v[2:3] offset:63488
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_f64__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: flat_load_b64 v[6:7], v[4:5]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB21_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_f64__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dwordx2 v[6:7], v[0:1]
+; GFX10-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB21_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_f64__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_atomic_add_f64 v[0:1], v[2:3] offset:63488
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_f64__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v8, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v9, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dwordx2 v[6:7], v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[8:9], v[4:7] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v7, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v6, v0
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB21_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_f64__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v9, vcc, -1, v1, vcc
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff804, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v7, v[0:1]
+; GFX8-NEXT: flat_load_dword v6, v[8:9]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[8:9], v[4:7] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v7, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB21_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_f64__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v8, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v9, vcc, -1, v1, vcc
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff804, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v7, v[0:1]
+; GFX7-NEXT: flat_load_dword v6, v[8:9]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_f64 v[4:5], v[6:7], v[2:3]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[8:9], v[4:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v7, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB21_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr %ptr, i64 -256
+ %unused = atomicrmw fadd ptr %gep, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; half
+; --------------------------------------------------------------------
+
+define half @flat_agent_atomic_fadd_ret_f16(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB22_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: flat_load_dword v4, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX940-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB22_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB22_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB22_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v4, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX90A-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB22_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v4
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX908-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX908-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX908-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB22_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB22_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v2, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_add_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v2, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB22_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr %ptr, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+define half @flat_agent_atomic_fadd_ret_f16__offset12b_pos(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB23_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[6:7], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v6
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: flat_load_dword v4, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX940-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB23_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB23_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB23_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v4, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX90A-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB23_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v4
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX908-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX908-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX908-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB23_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB23_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v2, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_add_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v2, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB23_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %result = atomicrmw fadd ptr %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+define half @flat_agent_atomic_fadd_ret_f16__offset12b_neg(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB24_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[6:7], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v6
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: flat_load_dword v4, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX940-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB24_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB24_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB24_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v4, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX90A-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB24_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v4
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX908-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX908-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX908-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB24_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB24_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v2, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_add_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v2, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB24_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 -1024
+ %result = atomicrmw fadd ptr %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+ }
+
+define void @flat_agent_atomic_fadd_noret_f16(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v4, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB25_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX940-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB25_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v4, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB25_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX10-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB25_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX90A-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB25_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX908-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB25_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX8-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB25_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: v_lshl_b32_e32 v3, 0xffff, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_not_b32_e32 v6, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v2, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX7-NEXT: v_add_f32_e32 v3, v3, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v2, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB25_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fadd ptr %ptr, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fadd_noret_f16__offset12b_pos(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v4, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB26_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX940-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB26_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v4, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB26_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX10-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB26_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX90A-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB26_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX908-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB26_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX8-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB26_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB26_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %unused = atomicrmw fadd ptr %gep, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fadd_noret_f16__offset12b_neg(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v4, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB27_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX940-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB27_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v4, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB27_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX10-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB27_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX90A-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB27_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX908-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB27_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX8-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB27_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB27_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 -1024
+ %unused = atomicrmw fadd ptr %gep, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define half @flat_agent_atomic_fadd_ret_f16__offset12b_pos__align4(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_pos__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f16_e32 v3, v4, v2
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB28_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_pos__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_add_f16_e32 v3, v5, v2
+; GFX940-NEXT: v_and_or_b32 v4, v5, s2, v3
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB28_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_pos__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f16_e32 v3, v4, v2
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB28_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_pos__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_add_f16_e32 v0, v1, v2
+; GFX10-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX10-NEXT: v_and_or_b32 v0, 0xffff0000, v1, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB28_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_pos__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_add_f16_e32 v3, v5, v2
+; GFX90A-NEXT: v_and_or_b32 v4, v5, s6, v3
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB28_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_pos__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_add_f16_e32 v3, v4, v2
+; GFX908-NEXT: v_and_or_b32 v3, v4, s6, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB28_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_pos__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_add_f16_e32 v0, v1, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX8-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB28_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_f16__offset12b_pos__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB28_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %result = atomicrmw fadd ptr %gep, half %val syncscope("agent") seq_cst, align 4
+ ret half %result
+}
+
+define void @flat_agent_atomic_fadd_noret_f16__offset12b__align4_pos(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b__align4_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v4, v[0:1] offset:2046
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_add_f16_e32 v3, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB29_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b__align4_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v5, v[0:1] offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_f16_e32 v3, v5, v2
+; GFX940-NEXT: v_and_or_b32 v4, v5, s2, v3
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB29_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b__align4_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v4, v[0:1] offset:2046
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_f16_e32 v3, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB29_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b__align4_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_f16_e32 v3, v4, v2
+; GFX10-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX10-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB29_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b__align4_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v5, v[0:1] offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_f16_e32 v3, v5, v2
+; GFX90A-NEXT: v_and_or_b32 v4, v5, s6, v3
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB29_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b__align4_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v4, v[0:1] offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_f16_e32 v3, v4, v2
+; GFX908-NEXT: v_and_or_b32 v3, v4, s6, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB29_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b__align4_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_f16_e32 v3, v4, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB29_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_f16__offset12b__align4_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v2
+; GFX7-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB29_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %unused = atomicrmw fadd ptr %gep, half %val syncscope("agent") seq_cst, align 4
+ ret void
+}
+
+define half @flat_system_atomic_fadd_ret_f16__offset12b_pos(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fadd_ret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB30_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fadd_ret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[6:7], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v6
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: flat_load_dword v4, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX940-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB30_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fadd_ret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB30_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fadd_ret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB30_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fadd_ret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v4, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX90A-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB30_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fadd_ret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v4
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX908-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX908-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX908-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB30_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fadd_ret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_add_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB30_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fadd_ret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v2, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_add_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v2, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB30_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %result = atomicrmw fadd ptr %gep, half %val seq_cst
+ ret half %result
+}
+
+define void @flat_system_atomic_fadd_noret_f16__offset12b_pos(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fadd_noret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v4, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB31_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fadd_noret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX940-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB31_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fadd_noret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v4, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB31_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fadd_noret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX10-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB31_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fadd_noret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX90A-NEXT: v_add_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB31_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fadd_noret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX908-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB31_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fadd_noret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX8-NEXT: v_add_f16_e32 v3, v3, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB31_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fadd_noret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB31_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %unused = atomicrmw fadd ptr %gep, half %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; bfloat
+; --------------------------------------------------------------------
+
+define bfloat @flat_agent_atomic_fadd_ret_bf16(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB32_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB32_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB32_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB32_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB32_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB32_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB32_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB32_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr %ptr, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+define bfloat @flat_agent_atomic_fadd_ret_bf16__offset12b_pos(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB33_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB33_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB33_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB33_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB33_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB33_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB33_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB33_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %result = atomicrmw fadd ptr %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+define bfloat @flat_agent_atomic_fadd_ret_bf16__offset12b_neg(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB34_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB34_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB34_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB34_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB34_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB34_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB34_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB34_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 -1024
+ %result = atomicrmw fadd ptr %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+ }
+
+define void @flat_agent_atomic_fadd_noret_bf16(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v4, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB35_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v4, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v4, v4, v2
+; GFX940-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX940-NEXT: v_add3_u32 v7, v7, v4, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v4, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB35_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v4, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB35_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB35_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v4, v4, v2
+; GFX90A-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX90A-NEXT: v_add3_u32 v7, v7, v4, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB35_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB35_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX8-NEXT: v_bfe_u32 v8, v3, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v3
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB35_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v3, 0xffff, v5
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v6, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB35_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fadd ptr %ptr, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fadd_noret_bf16__offset12b_pos(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB36_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB36_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB36_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB36_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB36_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB36_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB36_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB36_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %unused = atomicrmw fadd ptr %gep, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fadd_noret_bf16__offset12b_neg(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB37_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB37_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB37_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB37_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB37_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB37_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB37_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB37_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 -1024
+ %unused = atomicrmw fadd ptr %gep, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define bfloat @flat_agent_atomic_fadd_ret_bf16__offset12b_pos__align4(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_pos__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX12-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB38_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_pos__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX940-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX940-NEXT: v_bfe_u32 v4, v3, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX940-NEXT: v_add3_u32 v4, v4, v3, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v3, v4, v6, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX940-NEXT: v_and_or_b32 v4, v5, s3, v3
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB38_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_pos__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX11-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB38_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_pos__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX10-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX10-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX10-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX10-NEXT: v_and_or_b32 v5, 0xffff0000, v6, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB38_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_pos__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX90A-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX90A-NEXT: v_bfe_u32 v4, v3, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX90A-NEXT: v_add3_u32 v4, v4, v3, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e32 v3, v4, v6, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX90A-NEXT: v_and_or_b32 v4, v5, s7, v3
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB38_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_pos__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX908-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX908-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, s7, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB38_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_pos__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v0
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX8-NEXT: v_cndmask_b32_e32 v0, v5, v7, vcc
+; GFX8-NEXT: v_or_b32_sdwa v5, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB38_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_bf16__offset12b_pos__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX7-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB38_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %result = atomicrmw fadd ptr %gep, bfloat %val syncscope("agent") seq_cst, align 4
+ ret bfloat %result
+}
+
+define void @flat_agent_atomic_fadd_noret_bf16__offset12b__align4_pos(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b__align4_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB39_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b__align4_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v5, v5, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: v_and_or_b32 v2, v3, s3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB39_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b__align4_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB39_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b__align4_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB39_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b__align4_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v5, v5, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v3, s7, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB39_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b__align4_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v5, v5, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: v_and_or_b32 v2, v3, s7, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB39_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b__align4_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v2
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v6, v7, vcc
+; GFX8-NEXT: v_or_b32_sdwa v2, v5, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB39_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b__align4_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX7-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB39_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %unused = atomicrmw fadd ptr %gep, bfloat %val syncscope("agent") seq_cst, align 4
+ ret void
+}
+
+define bfloat @flat_system_atomic_fadd_ret_bf16__offset12b_pos(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB40_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB40_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB40_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB40_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB40_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB40_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB40_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fadd_ret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB40_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %result = atomicrmw fadd ptr %gep, bfloat %val seq_cst
+ ret bfloat %result
+}
+
+define void @flat_system_atomic_fadd_noret_bf16__offset12b_pos(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB41_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB41_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB41_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB41_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB41_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB41_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB41_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fadd_noret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_add_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB41_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %unused = atomicrmw fadd ptr %gep, bfloat %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x half>
+; --------------------------------------------------------------------
+
+define <2 x half> @flat_agent_atomic_fadd_ret_v2f16(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB42_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB42_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB42_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB42_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB42_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB42_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_add_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v5, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB42_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX7-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[6:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB42_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr %ptr, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define <2 x half> @flat_agent_atomic_fadd_ret_v2f16__offset12b_pos(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB43_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB43_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB43_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_pk_add_f16 v0, v1, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB43_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB43_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB43_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_add_f16_sdwa v0, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v5, v1, v2
+; GFX8-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB43_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v1, v[4:5]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v0
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[4:5], v[6:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB43_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 511
+ %result = atomicrmw fadd ptr %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define <2 x half> @flat_agent_atomic_fadd_ret_v2f16__offset12b_neg(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_v2f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB44_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_v2f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v4
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v5, vcc
+; GFX940-NEXT: flat_load_dword v0, v[0:1]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, v0
+; GFX940-NEXT: v_pk_add_f16 v0, v1, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB44_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_v2f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: flat_load_b32 v0, v[4:5]
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v1, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v0, v1, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v0, v[3:4], v[0:1] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB44_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_v2f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_pk_add_f16 v0, v1, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB44_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_v2f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v0, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: v_pk_add_f16 v0, v1, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB44_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_v2f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v0, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: v_pk_add_f16 v0, v1, v2
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB44_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_v2f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_add_f16_sdwa v0, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v5, v1, v2
+; GFX8-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB44_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_v2f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v1, v[4:5]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v0
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[4:5], v[6:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB44_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 -512
+ %result = atomicrmw fadd ptr %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define void @flat_agent_atomic_fadd_noret_v2f16(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v4, v[0:1]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB45_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB45_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v4, v[0:1]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB45_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB45_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB45_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB45_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v5, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB45_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: flat_atomic_cmpswap v7, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB45_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fadd ptr %ptr, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fadd_noret_v2f16__offset12b_pos(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v4, v[0:1] offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB46_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v5, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB46_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v4, v[0:1] offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB46_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB46_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v5, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB46_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v4, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB46_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v5, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB46_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: flat_atomic_cmpswap v7, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB46_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 511
+ %unused = atomicrmw fadd ptr %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fadd_noret_v2f16__offset12b_neg(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_v2f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v4, v[0:1] offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB47_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_v2f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX940-NEXT: flat_load_dword v5, v[4:5]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB47_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_v2f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: flat_load_b32 v4, v[3:4]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB47_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_v2f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB47_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_v2f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v1, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v0, v1, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB47_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_v2f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v1, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v0, v1, v2
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB47_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_v2f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v5, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB47_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_v2f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: flat_atomic_cmpswap v7, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB47_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 -512
+ %unused = atomicrmw fadd ptr %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x half> @flat_system_atomic_fadd_ret_v2f16__offset12b_pos(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB48_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB48_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB48_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_pk_add_f16 v0, v1, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB48_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB48_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB48_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_add_f16_sdwa v0, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v5, v1, v2
+; GFX8-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB48_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fadd_ret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v1, v[4:5]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v0
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[4:5], v[6:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB48_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 511
+ %result = atomicrmw fadd ptr %gep, <2 x half> %val seq_cst
+ ret <2 x half> %result
+}
+
+define void @flat_system_atomic_fadd_noret_v2f16__offset12b_pos(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v4, v[0:1] offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB49_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v5, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB49_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v4, v[0:1] offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB49_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB49_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v5, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB49_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v4, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB49_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v5, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB49_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fadd_noret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: flat_atomic_cmpswap v7, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB49_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 511
+ %unused = atomicrmw fadd ptr %gep, <2 x half> %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x bfloat>
+; --------------------------------------------------------------------
+
+define <2 x bfloat> @flat_agent_atomic_fadd_ret_v2bf16(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB50_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB50_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB50_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB50_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB50_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB50_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v3
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v3, 16
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB50_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v2, v2, v6, 16
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB50_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fadd ptr %ptr, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define <2 x bfloat> @flat_agent_atomic_fadd_ret_v2bf16__offset12b_pos(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB51_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
;
-; GFX940-LABEL: flat_agent_atomic_fadd_noret_v2f16_negoffset:
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_v2bf16__offset12b_pos:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
-; GFX940-NEXT: s_movk_i32 s0, 0xf800
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
-; GFX940-NEXT: flat_load_dword v5, v[4:5]
-; GFX940-NEXT: s_mov_b32 s1, -1
-; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
-; GFX940-NEXT: s_mov_b64 s[0:1], 0
-; GFX940-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB51_1: ; %atomicrmw.start
; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] offset:2044 sc0
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
-; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT: s_cbranch_execnz .LBB13_1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB51_1
; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX10-LABEL: flat_agent_atomic_fadd_noret_v2f16_negoffset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
-; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
-; GFX10-NEXT: s_mov_b32 s4, 0
-; GFX10-NEXT: flat_load_dword v4, v[0:1]
-; GFX10-NEXT: .LBB13_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX10-NEXT: v_mov_b32_e32 v4, v3
-; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_cbranch_execnz .LBB13_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: flat_agent_atomic_fadd_noret_v2f16_negoffset:
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_v2bf16__offset12b_pos:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
-; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
-; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
-; GFX11-NEXT: flat_load_b32 v4, v[3:4]
-; GFX11-NEXT: s_mov_b32 s0, 0
-; GFX11-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB51_1: ; %atomicrmw.start
; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_pk_add_f16 v3, v4, v2
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:2044 glc
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT: buffer_gl1_inv
; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX11-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX11-NEXT: s_cbranch_execnz .LBB13_1
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB51_1
; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
; GFX11-NEXT: s_setpc_b64 s[30:31]
;
-; GFX12-LABEL: flat_agent_atomic_fadd_noret_v2f16_negoffset:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: flat_load_b32 v4, v[0:1] offset:-2048
-; GFX12-NEXT: s_mov_b32 s0, 0
-; GFX12-NEXT: .LBB13_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: v_pk_add_f16 v3, v4, v2
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:-2048 th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
-; GFX12-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_cbranch_execnz .LBB13_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x half>, ptr %ptr, i64 -512
- %result = atomicrmw fadd ptr %gep, <2 x half> %val syncscope("agent") seq_cst
- ret void
-}
-
-define <2 x bfloat> @flat_agent_atomic_fadd_ret_v2bf16_offset(ptr %ptr, <2 x bfloat> %val) {
-; GFX7-LABEL: flat_agent_atomic_fadd_ret_v2bf16_offset:
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v0
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v0, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v0, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB51_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB51_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB51_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB51_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_v2bf16__offset12b_pos:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x3ffc, v0
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fc, v0
; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
; GFX7-NEXT: flat_load_dword v0, v[4:5]
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
@@ -3654,7 +12713,7 @@ define <2 x bfloat> @flat_agent_atomic_fadd_ret_v2bf16_offset(ptr %ptr, <2 x bfl
; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GFX7-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX7-NEXT: .LBB51_1: ; %atomicrmw.start
; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
@@ -3668,27 +12727,311 @@ define <2 x bfloat> @flat_agent_atomic_fadd_ret_v2bf16_offset(ptr %ptr, <2 x bfl
; GFX7-NEXT: v_alignbit_b32 v0, v0, v6, 16
; GFX7-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: buffer_wbinvl1_vol
+; GFX7-NEXT: buffer_wbinvl1
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB14_1
+; GFX7-NEXT: s_cbranch_execnz .LBB51_1
; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 511
+ %result = atomicrmw fadd ptr %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define <2 x bfloat> @flat_agent_atomic_fadd_ret_v2bf16__offset12b_neg(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_ret_v2bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB52_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_ret_v2bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v4
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v5, vcc
+; GFX940-NEXT: flat_load_dword v0, v[0:1]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v3, 0xffff0000, v7
+; GFX940-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX940-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX940-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v3, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v3
+; GFX940-NEXT: v_add3_u32 v6, v6, v0, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v3, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v0, v0
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v0, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v3, v0, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v0, v[4:5], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB52_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_ret_v2bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: flat_load_b32 v0, v[4:5]
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v0
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v0, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v0, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v0, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v0, v[3:4], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB52_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_ret_v2bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v0
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v0, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v0, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB52_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_ret_v2bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v0, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v3, 0xffff0000, v7
+; GFX90A-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX90A-NEXT: v_add_f32_e32 v3, v3, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v3, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v3
+; GFX90A-NEXT: v_add3_u32 v6, v6, v0, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v3, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX90A-NEXT: v_cndmask_b32_e64 v0, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v3, v0, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB52_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_ret_v2bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v0, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v0, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX908-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v0, s9
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB52_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
;
-; GFX8-LABEL: flat_agent_atomic_fadd_ret_v2bf16_offset:
+; GFX8-LABEL: flat_agent_atomic_fadd_ret_v2bf16__offset12b_neg:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x3ffc, v0
-; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
; GFX8-NEXT: flat_load_dword v0, v[3:4]
; GFX8-NEXT: s_mov_b64 s[6:7], 0
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX8-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX8-NEXT: .LBB52_1: ; %atomicrmw.start
; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_mov_b32_e32 v6, v0
@@ -3712,621 +13055,1102 @@ define <2 x bfloat> @flat_agent_atomic_fadd_ret_v2bf16_offset(ptr %ptr, <2 x bfl
; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: buffer_wbinvl1_vol
+; GFX8-NEXT: buffer_wbinvl1
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX8-NEXT: s_cbranch_execnz .LBB14_1
+; GFX8-NEXT: s_cbranch_execnz .LBB52_1
; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
-; GFX908-LABEL: flat_agent_atomic_fadd_ret_v2bf16_offset:
+; GFX7-LABEL: flat_agent_atomic_fadd_ret_v2bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[4:5]
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v0, v0, v6, 16
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB52_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 -512
+ %result = atomicrmw fadd ptr %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @flat_agent_atomic_fadd_noret_v2bf16(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB53_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB53_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB53_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB53_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB53_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_v2bf16:
; GFX908: ; %bb.0:
; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x3ffc, v0
-; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, 0, v1, vcc
-; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0x3000, v0
-; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX908-NEXT: flat_load_dword v0, v[0:1] offset:4092
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX908-NEXT: .LBB53_1: ; %atomicrmw.start
; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v6, v0
-; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v6
-; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX908-NEXT: v_add_f32_e32 v0, v0, v1
-; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX908-NEXT: v_bfe_u32 v7, v0, 16, 1
-; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v0
-; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX908-NEXT: v_add3_u32 v7, v7, v0, s8
-; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
-; GFX908-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX908-NEXT: v_perm_b32 v5, v5, v0, s9
-; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB14_1
+; GFX908-NEXT: s_cbranch_execnz .LBB53_1
; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
; GFX908-NEXT: s_setpc_b64 s[30:31]
;
-; GFX90A-LABEL: flat_agent_atomic_fadd_ret_v2bf16_offset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x3ffc, v0
-; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v1, vcc
-; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0x3000, v0
-; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX90A-NEXT: flat_load_dword v0, v[0:1] offset:4092
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 16, v2
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB14_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v7, v0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v7
-; GFX90A-NEXT: v_and_b32_e32 v3, 0xffff0000, v7
-; GFX90A-NEXT: v_add_f32_e32 v0, v0, v1
-; GFX90A-NEXT: v_add_f32_e32 v3, v3, v2
-; GFX90A-NEXT: v_bfe_u32 v6, v0, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v9, v3, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v0
-; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v3
-; GFX90A-NEXT: v_add3_u32 v6, v6, v0, s8
-; GFX90A-NEXT: v_add3_u32 v9, v9, v3, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
-; GFX90A-NEXT: v_cndmask_b32_e64 v0, v6, v8, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc
-; GFX90A-NEXT: v_perm_b32 v6, v3, v0, s9
-; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[6:7] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v7
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB14_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB53_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[4:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB53_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fadd ptr %ptr, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fadd_noret_v2bf16__offset12b_pos(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB54_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
;
-; GFX940-LABEL: flat_agent_atomic_fadd_ret_v2bf16_offset:
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_v2bf16__offset12b_pos:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v4, v0
-; GFX940-NEXT: v_mov_b32_e32 v5, v1
-; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0x3000, v4
-; GFX940-NEXT: s_mov_b64 s[0:1], 0x3ffc
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v5, vcc
-; GFX940-NEXT: flat_load_dword v0, v[0:1] offset:4092
-; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
; GFX940-NEXT: s_mov_b64 s[2:3], 0
-; GFX940-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX940-NEXT: .LBB54_1: ; %atomicrmw.start
; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v7, v0
-; GFX940-NEXT: v_lshlrev_b32_e32 v0, 16, v7
-; GFX940-NEXT: v_and_b32_e32 v3, 0xffff0000, v7
-; GFX940-NEXT: v_add_f32_e32 v0, v0, v1
-; GFX940-NEXT: v_add_f32_e32 v3, v3, v2
-; GFX940-NEXT: v_bfe_u32 v6, v0, 16, 1
-; GFX940-NEXT: v_bfe_u32 v9, v3, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v0
-; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v3
-; GFX940-NEXT: v_add3_u32 v6, v6, v0, s4
-; GFX940-NEXT: v_add3_u32 v9, v9, v3, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v0, v0
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v0, v6, v8, s[0:1]
-; GFX940-NEXT: v_perm_b32 v6, v3, v0, s5
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: flat_atomic_cmpswap v0, v[4:5], v[6:7] sc0
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 sc0
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v7
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB14_1
+; GFX940-NEXT: s_cbranch_execnz .LBB54_1
; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX10-LABEL: flat_agent_atomic_fadd_ret_v2bf16_offset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x3ffc, v0
-; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
-; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
-; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX10-NEXT: s_mov_b32 s5, 0
-; GFX10-NEXT: flat_load_dword v0, v[3:4]
-; GFX10-NEXT: .LBB14_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v6, v0
-; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
-; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX10-NEXT: v_add_f32_e32 v0, v0, v1
-; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX10-NEXT: v_bfe_u32 v7, v0, 16, 1
-; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v0
-; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX10-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
-; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX10-NEXT: v_cmp_u_f32_e64 s4, v0, v0
-; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX10-NEXT: v_cndmask_b32_e64 v0, v7, v9, s4
-; GFX10-NEXT: v_perm_b32 v5, v5, v0, 0x7060302
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
-; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_cbranch_execnz .LBB14_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: flat_agent_atomic_fadd_ret_v2bf16_offset:
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_v2bf16__offset12b_pos:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
; GFX11-NEXT: s_mov_b32 s1, 0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x3000, v3
-; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, 0, v1, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x3ffc, v3
-; GFX11-NEXT: flat_load_b32 v0, v[4:5] offset:4092
-; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX11-NEXT: .LBB54_1: ; %atomicrmw.start
; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v6, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v7, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v0
-; GFX11-NEXT: v_cmp_u_f32_e64 s0, v0, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX11-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_cndmask_b32_e64 v0, v7, v9, s0
-; GFX11-NEXT: v_perm_b32 v5, v5, v0, 0x7060302
+; GFX11-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: flat_atomic_cmpswap_b32 v0, v[3:4], v[5:6] glc
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 glc
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT: buffer_gl1_inv
; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_cbranch_execnz .LBB14_1
+; GFX11-NEXT: s_cbranch_execnz .LBB54_1
; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
; GFX11-NEXT: s_setpc_b64 s[30:31]
;
-; GFX12-LABEL: flat_agent_atomic_fadd_ret_v2bf16_offset:
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB54_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB54_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB54_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB54_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[4:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB54_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 511
+ %unused = atomicrmw fadd ptr %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fadd_noret_v2bf16__offset12b_neg(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fadd_noret_v2bf16__offset12b_neg:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT: s_wait_expcnt 0x0
; GFX12-NEXT: s_wait_samplecnt 0x0
; GFX12-NEXT: s_wait_bvhcnt 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:16380
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX12-NEXT: .LBB55_1: ; %atomicrmw.start
; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: v_mov_b32_e32 v6, v3
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX12-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX12-NEXT: v_add_f32_e32 v3, v3, v4
-; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
-; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
-; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:16380 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:-2048 th:TH_ATOMIC_RETURN
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB14_1
+; GFX12-NEXT: s_cbranch_execnz .LBB55_1
; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: v_mov_b32_e32 v0, v3
; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 4095
- %result = atomicrmw fadd ptr %gep, <2 x bfloat> %val syncscope("agent") seq_cst
- ret <2 x bfloat> %result
-}
-
-define <2 x bfloat> @flat_agent_atomic_fadd_ret_v2bf16_negoffset(ptr %ptr, <2 x bfloat> %val) {
-; GFX7-LABEL: flat_agent_atomic_fadd_ret_v2bf16_negoffset:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
-; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
-; GFX7-NEXT: flat_load_dword v0, v[4:5]
-; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
-; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v1
-; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
-; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GFX7-NEXT: .LBB15_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
-; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
-; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v1
-; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v0
-; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
-; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
-; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
-; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v7
-; GFX7-NEXT: v_alignbit_b32 v0, v0, v6, 16
-; GFX7-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: buffer_wbinvl1_vol
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB15_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: flat_agent_atomic_fadd_ret_v2bf16_negoffset:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
-; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
-; GFX8-NEXT: flat_load_dword v0, v[3:4]
-; GFX8-NEXT: s_mov_b64 s[6:7], 0
-; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
-; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX8-NEXT: .LBB15_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_mov_b32_e32 v6, v0
-; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
-; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX8-NEXT: v_add_f32_e32 v0, v0, v1
-; GFX8-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
-; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
-; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
-; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
-; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
-; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
-; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
-; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
-; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
-; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
-; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: buffer_wbinvl1_vol
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
-; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX8-NEXT: s_cbranch_execnz .LBB15_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX8-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: flat_agent_atomic_fadd_ret_v2bf16_negoffset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
-; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
-; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
-; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
-; GFX908-NEXT: flat_load_dword v0, v[0:1]
-; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v1, 16, v2
-; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB15_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v6, v0
-; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v6
-; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX908-NEXT: v_add_f32_e32 v0, v0, v1
-; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX908-NEXT: v_bfe_u32 v7, v0, 16, 1
-; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v0
-; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX908-NEXT: v_add3_u32 v7, v7, v0, s8
-; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
-; GFX908-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
-; GFX908-NEXT: v_perm_b32 v5, v5, v0, s9
-; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
-; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB15_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: flat_agent_atomic_fadd_ret_v2bf16_negoffset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
-; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
-; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
-; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
-; GFX90A-NEXT: flat_load_dword v0, v[0:1]
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 16, v2
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB15_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v7, v0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v7
-; GFX90A-NEXT: v_and_b32_e32 v3, 0xffff0000, v7
-; GFX90A-NEXT: v_add_f32_e32 v0, v0, v1
-; GFX90A-NEXT: v_add_f32_e32 v3, v3, v2
-; GFX90A-NEXT: v_bfe_u32 v6, v0, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v9, v3, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v0
-; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v3
-; GFX90A-NEXT: v_add3_u32 v6, v6, v0, s8
-; GFX90A-NEXT: v_add3_u32 v9, v9, v3, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
-; GFX90A-NEXT: v_cndmask_b32_e64 v0, v6, v8, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc
-; GFX90A-NEXT: v_perm_b32 v6, v3, v0, s9
-; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[6:7] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v7
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB15_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
-; GFX940-LABEL: flat_agent_atomic_fadd_ret_v2bf16_negoffset:
+; GFX940-LABEL: flat_agent_atomic_fadd_noret_v2bf16__offset12b_neg:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v4, v0
-; GFX940-NEXT: v_mov_b32_e32 v5, v1
-; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v4
+; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
; GFX940-NEXT: s_movk_i32 s0, 0xf800
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v5, vcc
-; GFX940-NEXT: flat_load_dword v0, v[0:1]
+; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX940-NEXT: flat_load_dword v3, v[4:5]
; GFX940-NEXT: s_mov_b32 s1, -1
-; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
+; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
; GFX940-NEXT: s_mov_b64 s[2:3], 0
-; GFX940-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB15_1: ; %atomicrmw.start
-; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v7, v0
-; GFX940-NEXT: v_lshlrev_b32_e32 v0, 16, v7
-; GFX940-NEXT: v_and_b32_e32 v3, 0xffff0000, v7
-; GFX940-NEXT: v_add_f32_e32 v0, v0, v1
-; GFX940-NEXT: v_add_f32_e32 v3, v3, v2
-; GFX940-NEXT: v_bfe_u32 v6, v0, 16, 1
-; GFX940-NEXT: v_bfe_u32 v9, v3, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v0
-; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v3
-; GFX940-NEXT: v_add3_u32 v6, v6, v0, s4
-; GFX940-NEXT: v_add3_u32 v9, v9, v3, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v0, v0
+; GFX940-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v0, v6, v8, s[0:1]
-; GFX940-NEXT: v_perm_b32 v6, v3, v0, s5
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: flat_atomic_cmpswap v0, v[4:5], v[6:7] sc0
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v7
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB15_1
+; GFX940-NEXT: s_cbranch_execnz .LBB55_1
; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX10-LABEL: flat_agent_atomic_fadd_ret_v2bf16_negoffset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
-; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
-; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
-; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX10-NEXT: s_mov_b32 s5, 0
-; GFX10-NEXT: flat_load_dword v0, v[3:4]
-; GFX10-NEXT: .LBB15_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_mov_b32_e32 v6, v0
-; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
-; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX10-NEXT: v_add_f32_e32 v0, v0, v1
-; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX10-NEXT: v_bfe_u32 v7, v0, 16, 1
-; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v0
-; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX10-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
-; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX10-NEXT: v_cmp_u_f32_e64 s4, v0, v0
-; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX10-NEXT: v_cndmask_b32_e64 v0, v7, v9, s4
-; GFX10-NEXT: v_perm_b32 v5, v5, v0, 0x7060302
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
-; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_cbranch_execnz .LBB15_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: flat_agent_atomic_fadd_ret_v2bf16_negoffset:
+; GFX11-LABEL: flat_agent_atomic_fadd_noret_v2bf16__offset12b_neg:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v3, v0
-; GFX11-NEXT: s_mov_b32 s1, 0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v3
-; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v3
-; GFX11-NEXT: flat_load_b32 v0, v[4:5]
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
-; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: flat_load_b32 v3, v[3:4]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX11-NEXT: .LBB55_1: ; %atomicrmw.start
; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_mov_b32_e32 v6, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
-; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
-; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
-; GFX11-NEXT: v_add_f32_e32 v0, v0, v1
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
-; GFX11-NEXT: v_bfe_u32 v7, v0, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v0
-; GFX11-NEXT: v_cmp_u_f32_e64 s0, v0, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
-; GFX11-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_cndmask_b32_e64 v0, v7, v9, s0
-; GFX11-NEXT: v_perm_b32 v5, v5, v0, 0x7060302
+; GFX11-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: flat_atomic_cmpswap_b32 v0, v[3:4], v[5:6] glc
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT: buffer_gl1_inv
; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_cbranch_execnz .LBB15_1
+; GFX11-NEXT: s_cbranch_execnz .LBB55_1
; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
; GFX11-NEXT: s_setpc_b64 s[30:31]
;
-; GFX12-LABEL: flat_agent_atomic_fadd_ret_v2bf16_negoffset:
+; GFX10-LABEL: flat_agent_atomic_fadd_noret_v2bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB55_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fadd_noret_v2bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v1, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX90A-NEXT: v_add_f32_e32 v0, v0, v3
+; GFX90A-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX90A-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v0, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX90A-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v0, v6, v0, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB55_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fadd_noret_v2bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v1, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX908-NEXT: v_add_f32_e32 v0, v0, v5
+; GFX908-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX908-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v0, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX908-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v0, v6, v0, s9
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB55_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fadd_noret_v2bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB55_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fadd_noret_v2bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[4:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB55_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 -512
+ %unused = atomicrmw fadd ptr %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x bfloat> @flat_system_atomic_fadd_ret_v2bf16__offset12b_pos(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fadd_ret_v2bf16__offset12b_pos:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT: s_wait_expcnt 0x0
; GFX12-NEXT: s_wait_samplecnt 0x0
; GFX12-NEXT: s_wait_bvhcnt 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX12-NEXT: .LBB56_1: ; %atomicrmw.start
; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT: v_mov_b32_e32 v6, v3
@@ -4351,341 +14175,331 @@ define <2 x bfloat> @flat_agent_atomic_fadd_ret_v2bf16_negoffset(ptr %ptr, <2 x
; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:2044 th:TH_ATOMIC_RETURN
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB15_1
+; GFX12-NEXT: s_cbranch_execnz .LBB56_1
; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
; GFX12-NEXT: v_mov_b32_e32 v0, v3
; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 -512
- %result = atomicrmw fadd ptr %gep, <2 x bfloat> %val syncscope("agent") seq_cst
- ret <2 x bfloat> %result
-}
-
-define void @flat_agent_atomic_fadd_noret_v2bf16_offset(ptr %ptr, <2 x bfloat> %val) {
-; GFX7-LABEL: flat_agent_atomic_fadd_noret_v2bf16_offset:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x3ffc, v0
-; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GFX7-NEXT: flat_load_dword v5, v[0:1]
-; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
-; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX7-NEXT: .LBB16_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
-; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
-; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
-; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
-; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
-; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX7-NEXT: v_alignbit_b32 v5, v4, v5, 16
-; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v7
-; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
-; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[4:5] glc
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: buffer_wbinvl1_vol
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
-; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v6
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB16_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: flat_agent_atomic_fadd_noret_v2bf16_offset:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x3ffc, v0
-; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GFX8-NEXT: flat_load_dword v3, v[0:1]
-; GFX8-NEXT: s_mov_b64 s[6:7], 0
-; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX8-NEXT: .LBB16_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX8-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX8-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
-; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
-; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
-; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
-; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
-; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: buffer_wbinvl1_vol
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX8-NEXT: v_mov_b32_e32 v3, v2
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX8-NEXT: s_cbranch_execnz .LBB16_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX8-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: flat_agent_atomic_fadd_noret_v2bf16_offset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x3ffc, v0
-; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, 0, v1, vcc
-; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0x3000, v0
-; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX908-NEXT: flat_load_dword v1, v[0:1] offset:4092
-; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v5, 16, v2
-; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB16_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v1
-; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
-; GFX908-NEXT: v_add_f32_e32 v0, v0, v5
-; GFX908-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX908-NEXT: v_bfe_u32 v7, v0, 16, 1
-; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v0
-; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX908-NEXT: v_add3_u32 v7, v7, v0, s8
-; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
-; GFX908-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX908-NEXT: v_perm_b32 v0, v6, v0, s9
-; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX908-NEXT: v_mov_b32_e32 v1, v0
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB16_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: flat_agent_atomic_fadd_noret_v2bf16_offset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x3ffc, v0
-; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v1, vcc
-; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0x3000, v0
-; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
-; GFX90A-NEXT: flat_load_dword v1, v[0:1] offset:4092
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v2
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB16_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v1
-; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
-; GFX90A-NEXT: v_add_f32_e32 v0, v0, v3
-; GFX90A-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX90A-NEXT: v_bfe_u32 v7, v0, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v0
-; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX90A-NEXT: v_add3_u32 v7, v7, v0, s8
-; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
-; GFX90A-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX90A-NEXT: v_perm_b32 v0, v6, v0, s9
-; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: v_mov_b32_e32 v1, v0
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB16_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
-; GFX940-LABEL: flat_agent_atomic_fadd_noret_v2bf16_offset:
+; GFX940-LABEL: flat_system_atomic_fadd_ret_v2bf16__offset12b_pos:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0x3000, v0
-; GFX940-NEXT: s_mov_b64 s[0:1], 0x3ffc
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v1, vcc
-; GFX940-NEXT: flat_load_dword v3, v[4:5] offset:4092
-; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
; GFX940-NEXT: s_mov_b64 s[2:3], 0
; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX940-NEXT: s_movk_i32 s4, 0x7fff
-; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX940-NEXT: .LBB56_1: ; %atomicrmw.start
; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX940-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX940-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
-; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
-; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] offset:2044 sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB16_1
+; GFX940-NEXT: s_cbranch_execnz .LBB56_1
; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX10-LABEL: flat_agent_atomic_fadd_noret_v2bf16_offset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x3ffc, v0
-; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
-; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX10-NEXT: s_mov_b32 s5, 0
-; GFX10-NEXT: flat_load_dword v3, v[0:1]
-; GFX10-NEXT: .LBB16_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX10-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX10-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
-; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
-; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX10-NEXT: v_mov_b32_e32 v3, v2
-; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_cbranch_execnz .LBB16_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: flat_agent_atomic_fadd_noret_v2bf16_offset:
+; GFX11-LABEL: flat_system_atomic_fadd_ret_v2bf16__offset12b_pos:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x3000, v0
-; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0x3ffc, v0
-; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
-; GFX11-NEXT: flat_load_b32 v3, v[3:4] offset:4092
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-NEXT: s_mov_b32 s1, 0
; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX11-NEXT: .LBB56_1: ; %atomicrmw.start
; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX11-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:2044 glc
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT: buffer_gl1_inv
; GFX11-NEXT: buffer_gl0_inv
-; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_cbranch_execnz .LBB16_1
+; GFX11-NEXT: s_cbranch_execnz .LBB56_1
; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
; GFX11-NEXT: s_setpc_b64 s[30:31]
;
-; GFX12-LABEL: flat_agent_atomic_fadd_noret_v2bf16_offset:
+; GFX10-LABEL: flat_system_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v0
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v0, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v0, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB56_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB56_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_add_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB56_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB56_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fadd_ret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[4:5]
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v0, v0, v6, 16
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB56_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 511
+ %result = atomicrmw fadd ptr %gep, <2 x bfloat> %val seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @flat_system_atomic_fadd_noret_v2bf16__offset12b_pos(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fadd_noret_v2bf16__offset12b_pos:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT: s_wait_expcnt 0x0
; GFX12-NEXT: s_wait_samplecnt 0x0
; GFX12-NEXT: s_wait_bvhcnt 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:16380
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX12-NEXT: .LBB57_1: ; %atomicrmw.start
; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
@@ -4708,293 +14522,70 @@ define void @flat_agent_atomic_fadd_noret_v2bf16_offset(ptr %ptr, <2 x bfloat> %
; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16380 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 th:TH_ATOMIC_RETURN
; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
; GFX12-NEXT: v_mov_b32_e32 v3, v2
; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB16_1
+; GFX12-NEXT: s_cbranch_execnz .LBB57_1
; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 4095
- %result = atomicrmw fadd ptr %gep, <2 x bfloat> %val syncscope("agent") seq_cst
- ret void
-}
-
-define void @flat_agent_atomic_fadd_noret_v2bf16_negoffset(ptr %ptr, <2 x bfloat> %val) {
-; GFX7-LABEL: flat_agent_atomic_fadd_noret_v2bf16_negoffset:
-; GFX7: ; %bb.0:
-; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
-; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
-; GFX7-NEXT: flat_load_dword v5, v[0:1]
-; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
-; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
-; GFX7-NEXT: s_mov_b64 s[4:5], 0
-; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
-; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX7-NEXT: .LBB17_1: ; %atomicrmw.start
-; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
-; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
-; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
-; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
-; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
-; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
-; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX7-NEXT: v_alignbit_b32 v5, v4, v5, 16
-; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v7
-; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
-; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[4:5] glc
-; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX7-NEXT: buffer_wbinvl1_vol
-; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
-; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
-; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v6
-; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_cbranch_execnz .LBB17_1
-; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX7-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX8-LABEL: flat_agent_atomic_fadd_noret_v2bf16_negoffset:
-; GFX8: ; %bb.0:
-; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
-; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
-; GFX8-NEXT: flat_load_dword v3, v[0:1]
-; GFX8-NEXT: s_mov_b64 s[6:7], 0
-; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX8-NEXT: .LBB17_1: ; %atomicrmw.start
-; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX8-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX8-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
-; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
-; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
-; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
-; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
-; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
-; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
-; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
-; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
-; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX8-NEXT: buffer_wbinvl1_vol
-; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX8-NEXT: v_mov_b32_e32 v3, v2
-; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX8-NEXT: s_cbranch_execnz .LBB17_1
-; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX8-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: flat_agent_atomic_fadd_noret_v2bf16_negoffset:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
-; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
-; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
-; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
-; GFX908-NEXT: flat_load_dword v1, v[0:1]
-; GFX908-NEXT: s_mov_b64 s[6:7], 0
-; GFX908-NEXT: v_lshlrev_b32_e32 v5, 16, v2
-; GFX908-NEXT: s_movk_i32 s8, 0x7fff
-; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX908-NEXT: s_mov_b32 s9, 0x7060302
-; GFX908-NEXT: .LBB17_1: ; %atomicrmw.start
-; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v1
-; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
-; GFX908-NEXT: v_add_f32_e32 v0, v0, v5
-; GFX908-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX908-NEXT: v_bfe_u32 v7, v0, 16, 1
-; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v0
-; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX908-NEXT: v_add3_u32 v7, v7, v0, s8
-; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
-; GFX908-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
-; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX908-NEXT: v_perm_b32 v0, v6, v0, s9
-; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: buffer_wbinvl1_vol
-; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX908-NEXT: v_mov_b32_e32 v1, v0
-; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_cbranch_execnz .LBB17_1
-; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX908-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX90A-LABEL: flat_agent_atomic_fadd_noret_v2bf16_negoffset:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
-; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
-; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
-; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
-; GFX90A-NEXT: flat_load_dword v1, v[0:1]
-; GFX90A-NEXT: s_mov_b64 s[6:7], 0
-; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v2
-; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
-; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
-; GFX90A-NEXT: .LBB17_1: ; %atomicrmw.start
-; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v1
-; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
-; GFX90A-NEXT: v_add_f32_e32 v0, v0, v3
-; GFX90A-NEXT: v_add_f32_e32 v6, v6, v2
-; GFX90A-NEXT: v_bfe_u32 v7, v0, 16, 1
-; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v0
-; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX90A-NEXT: v_add3_u32 v7, v7, v0, s8
-; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
-; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
-; GFX90A-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
-; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX90A-NEXT: v_perm_b32 v0, v6, v0, s9
-; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
-; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: buffer_wbinvl1_vol
-; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
-; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
-; GFX90A-NEXT: v_mov_b32_e32 v1, v0
-; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_cbranch_execnz .LBB17_1
-; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
-; GFX940-LABEL: flat_agent_atomic_fadd_noret_v2bf16_negoffset:
+; GFX940-LABEL: flat_system_atomic_fadd_noret_v2bf16__offset12b_pos:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
-; GFX940-NEXT: s_movk_i32 s0, 0xf800
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
-; GFX940-NEXT: flat_load_dword v3, v[4:5]
-; GFX940-NEXT: s_mov_b32 s1, -1
-; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
; GFX940-NEXT: s_mov_b64 s[2:3], 0
; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX940-NEXT: s_movk_i32 s4, 0x7fff
; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
; GFX940-NEXT: s_mov_b32 s5, 0x7060302
-; GFX940-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX940-NEXT: .LBB57_1: ; %atomicrmw.start
; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX940-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX940-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
-; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
-; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
-; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
-; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
-; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
-; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
-; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
-; GFX940-NEXT: buffer_wbl2 sc1
-; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
-; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT: buffer_inv sc1
-; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_cbranch_execnz .LBB17_1
-; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX10-LABEL: flat_agent_atomic_fadd_noret_v2bf16_negoffset:
-; GFX10: ; %bb.0:
-; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
-; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
-; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX10-NEXT: s_mov_b32 s5, 0
-; GFX10-NEXT: flat_load_dword v3, v[0:1]
-; GFX10-NEXT: .LBB17_1: ; %atomicrmw.start
-; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX10-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX10-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
-; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
-; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX10-NEXT: v_mov_b32_e32 v3, v2
-; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
-; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_cbranch_execnz .LBB17_1
-; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
-; GFX10-NEXT: s_setpc_b64 s[30:31]
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB57_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
;
-; GFX11-LABEL: flat_agent_atomic_fadd_noret_v2bf16_negoffset:
+; GFX11-LABEL: flat_system_atomic_fadd_noret_v2bf16__offset12b_pos:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
-; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
-; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
-; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
-; GFX11-NEXT: flat_load_b32 v3, v[3:4]
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
; GFX11-NEXT: s_mov_b32 s1, 0
; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
; GFX11-NEXT: .p2align 6
-; GFX11-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX11-NEXT: .LBB57_1: ; %atomicrmw.start
; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
@@ -5017,7 +14608,7 @@ define void @flat_agent_atomic_fadd_noret_v2bf16_negoffset(ptr %ptr, <2 x bfloat
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 glc
; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT: buffer_gl1_inv
; GFX11-NEXT: buffer_gl0_inv
@@ -5026,59 +14617,217 @@ define void @flat_agent_atomic_fadd_noret_v2bf16_negoffset(ptr %ptr, <2 x bfloat
; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX11-NEXT: s_cbranch_execnz .LBB17_1
+; GFX11-NEXT: s_cbranch_execnz .LBB57_1
; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
; GFX11-NEXT: s_setpc_b64 s[30:31]
;
-; GFX12-LABEL: flat_agent_atomic_fadd_noret_v2bf16_negoffset:
-; GFX12: ; %bb.0:
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: s_wait_expcnt 0x0
-; GFX12-NEXT: s_wait_samplecnt 0x0
-; GFX12-NEXT: s_wait_bvhcnt 0x0
-; GFX12-NEXT: s_wait_kmcnt 0x0
-; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
-; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
-; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
-; GFX12-NEXT: s_mov_b32 s1, 0
-; GFX12-NEXT: .LBB17_1: ; %atomicrmw.start
-; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
-; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-NEXT: v_add_f32_e32 v2, v2, v4
-; GFX12-NEXT: v_add_f32_e32 v6, v6, v5
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
-; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
-; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
-; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
-; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
-; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
-; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
-; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
-; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
-; GFX12-NEXT: s_wait_storecnt 0x0
-; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:-2048 th:TH_ATOMIC_RETURN
-; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT: global_inv scope:SCOPE_DEV
-; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX12-NEXT: v_mov_b32_e32 v3, v2
-; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_cbranch_execnz .LBB17_1
-; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT: s_setpc_b64 s[30:31]
- %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 -512
- %result = atomicrmw fadd ptr %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+; GFX10-LABEL: flat_system_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB57_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB57_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB57_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_add_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_add_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB57_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fadd_noret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[4:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB57_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 511
+ %unused = atomicrmw fadd ptr %gep, <2 x bfloat> %val seq_cst
ret void
}
+
+attributes #0 = { nounwind "amdgpu-unsafe-fp-atomics"="true" }
+attributes #1 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" "amdgpu-unsafe-fp-atomics"="true" }
diff --git a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmax.ll b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmax.ll
new file mode 100644
index 0000000000000..00d418cd3912c
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmax.ll
@@ -0,0 +1,15756 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 < %s | FileCheck -check-prefix=GFX12 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx940 < %s | FileCheck -check-prefix=GFX940 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx90a < %s | FileCheck -check-prefix=GFX90A %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx908 < %s | FileCheck -check-prefix=GFX908 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tonga < %s | FileCheck -check-prefix=GFX8 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=hawaii < %s | FileCheck -check-prefix=GFX7 %s
+
+; --------------------------------------------------------------------
+; float
+; --------------------------------------------------------------------
+
+define float @flat_agent_atomic_fmax_ret_f32(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB0_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB0_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB0_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX10-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB0_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB0_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB0_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX8-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_mul_f32_e32 v3, 1.0, v4
+; GFX8-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB0_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v4
+; GFX7-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB0_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr %ptr, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @flat_agent_atomic_fmax_ret_f32__offset12b_pos(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB1_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB1_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB1_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_max_f32_e32 v0, v6, v6
+; GFX10-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB1_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB1_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB1_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB1_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX7-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB1_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %result = atomicrmw fmax ptr %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @flat_agent_atomic_fmax_ret_f32__offset12b_neg(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB2_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v4
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v5, vcc
+; GFX940-NEXT: flat_load_dword v0, v[0:1]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX940-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_max_f32_e32 v0, v3, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v0, v1
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v0, v[4:5], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB2_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: flat_load_b32 v0, v[4:5]
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX11-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v0, v6, v6
+; GFX11-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v0, v[3:4], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB2_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_max_f32_e32 v0, v6, v6
+; GFX10-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB2_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v0, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX90A-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_max_f32_e32 v0, v3, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v0, v1
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB2_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v0, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX908-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v0
+; GFX908-NEXT: v_max_f32_e32 v0, v6, v6
+; GFX908-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB2_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB2_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX7-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 -512
+ %result = atomicrmw fmax ptr %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define void @flat_agent_atomic_fmax_noret_f32(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB3_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB3_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB3_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB3_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB3_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB3_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB3_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB3_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmax ptr %ptr, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmax_noret_f32__offset12b_pos(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB4_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB4_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB4_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB4_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB4_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB4_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB4_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB4_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %unused = atomicrmw fmax ptr %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmax_noret_f32__offset12b_neg(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB5_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX940-NEXT: flat_load_dword v3, v[4:5]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB5_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: flat_load_b32 v3, v[3:4]
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB5_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB5_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v1, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v0, v1, v1
+; GFX90A-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB5_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v1, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v0, v1, v1
+; GFX908-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 -512
+ %unused = atomicrmw fmax ptr %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define float @flat_system_atomic_fmax_ret_f32__offset12b_pos(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fmax_ret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB6_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmax_ret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB6_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmax_ret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB6_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmax_ret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_max_f32_e32 v0, v6, v6
+; GFX10-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB6_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmax_ret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB6_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmax_ret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB6_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmax_ret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB6_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmax_ret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX7-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB6_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %result = atomicrmw fmax ptr %gep, float %val seq_cst
+ ret float %result
+}
+
+define void @flat_system_atomic_fmax_noret_f32__offset12b_pos(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fmax_noret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB7_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmax_noret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB7_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmax_noret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB7_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmax_noret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB7_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmax_noret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB7_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmax_noret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB7_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmax_noret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB7_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmax_noret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB7_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %unused = atomicrmw fmax ptr %gep, float %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; float with ftz/daz
+; --------------------------------------------------------------------
+
+define float @flat_agent_atomic_fmax_ret_f32__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_f32__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB8_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_f32__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB8_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_f32__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB8_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_f32__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX10-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB8_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_f32__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB8_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_f32__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_f32__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX8-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_mul_f32_e32 v3, 1.0, v4
+; GFX8-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_f32__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v4
+; GFX7-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB8_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr %ptr, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @flat_agent_atomic_fmax_ret_f32__offset12b_pos__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB9_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB9_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB9_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_max_f32_e32 v0, v6, v6
+; GFX10-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB9_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB9_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB9_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB9_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX7-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB9_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %result = atomicrmw fmax ptr %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @flat_agent_atomic_fmax_ret_f32__offset12b_neg__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_neg__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB10_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_neg__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v4
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v5, vcc
+; GFX940-NEXT: flat_load_dword v0, v[0:1]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX940-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_max_f32_e32 v0, v3, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v0, v1
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v0, v[4:5], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB10_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_neg__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: flat_load_b32 v0, v[4:5]
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX11-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v0, v6, v6
+; GFX11-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v0, v[3:4], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB10_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_neg__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_max_f32_e32 v0, v6, v6
+; GFX10-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB10_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_neg__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v0, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX90A-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_max_f32_e32 v0, v3, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v0, v1
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB10_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_neg__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v0, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX908-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v0
+; GFX908-NEXT: v_max_f32_e32 v0, v6, v6
+; GFX908-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB10_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_neg__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB10_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_f32__offset12b_neg__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX7-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB10_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 -512
+ %result = atomicrmw fmax ptr %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define void @flat_agent_atomic_fmax_noret_f32__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_f32__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB11_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_f32__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB11_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_f32__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB11_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_f32__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB11_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_f32__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_f32__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_f32__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_f32__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB11_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmax ptr %ptr, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmax_noret_f32__offset12b_pos__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB12_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB12_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB12_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB12_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB12_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB12_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB12_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB12_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %unused = atomicrmw fmax ptr %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmax_noret_f32__offset12b_neg__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_neg__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB13_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_neg__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX940-NEXT: flat_load_dword v3, v[4:5]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB13_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_neg__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: flat_load_b32 v3, v[3:4]
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB13_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_neg__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB13_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_neg__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v1, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v0, v1, v1
+; GFX90A-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB13_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_neg__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v1, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v0, v1, v1
+; GFX908-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB13_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_neg__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB13_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_f32__offset12b_neg__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB13_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 -512
+ %unused = atomicrmw fmax ptr %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define float @flat_system_atomic_fmax_ret_f32__offset12b_pos__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_system_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB14_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB14_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB14_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_max_f32_e32 v0, v6, v6
+; GFX10-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB14_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_max_f32_e32 v4, v3, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmax_ret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX7-NEXT: v_max_f32_e32 v5, v0, v1
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %result = atomicrmw fmax ptr %gep, float %val seq_cst
+ ret float %result
+}
+
+define void @flat_system_atomic_fmax_noret_f32__offset12b_pos__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_system_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB15_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB15_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB15_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB15_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB15_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB15_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB15_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmax_noret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB15_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %unused = atomicrmw fmax ptr %gep, float %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; double
+; --------------------------------------------------------------------
+
+define double @flat_agent_atomic_fmax_ret_f64(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b64 v[4:5], v[0:1]
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[6:7], v[6:7]
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[4:5], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB16_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX940-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB16_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b64 v[4:5], v[0:1]
+; GFX11-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX11-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB16_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX10-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX10-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB16_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX90A-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX90A-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB16_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX908-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX908-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB16_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 4, v0
+; GFX8-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: flat_load_dword v5, v[5:6]
+; GFX8-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v7, v5
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX8-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB16_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: v_mov_b32_e32 v1, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v5, vcc, 4, v0
+; GFX7-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: flat_load_dword v5, v[5:6]
+; GFX7-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX7-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB16_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v4
+; GFX7-NEXT: v_mov_b32_e32 v1, v5
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr %ptr, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define double @flat_agent_atomic_fmax_ret_f64__offset12b_pos(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_f64__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b64 v[4:5], v[0:1] offset:2040
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[6:7], v[6:7]
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[4:5], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] offset:2040 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB17_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_f64__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dwordx2 v[4:5], v[0:1] offset:2040
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX940-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] offset:2040 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB17_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_f64__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b64 v[4:5], v[0:1] offset:2040
+; GFX11-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX11-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] offset:2040 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB17_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_f64__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7f8, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dwordx2 v[0:1], v[4:5]
+; GFX10-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v9, v1
+; GFX10-NEXT: v_mov_b32_e32 v8, v0
+; GFX10-NEXT: v_max_f64 v[0:1], v[8:9], v[8:9]
+; GFX10-NEXT: v_max_f64 v[6:7], v[0:1], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[8:9]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB17_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_f64__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dwordx2 v[4:5], v[0:1] offset:2040
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX90A-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX90A-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] offset:2040 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB17_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_f64__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dwordx2 v[4:5], v[0:1] offset:2040
+; GFX908-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX908-NEXT: v_max_f64 v[4:5], v[4:5], v[2:3]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] offset:2040 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_f64__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7f8, v0
+; GFX8-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v1, v[0:1]
+; GFX8-NEXT: flat_load_dword v0, v[4:5]
+; GFX8-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v9, v1
+; GFX8-NEXT: v_mov_b32_e32 v8, v0
+; GFX8-NEXT: v_max_f64 v[0:1], v[8:9], v[8:9]
+; GFX8-NEXT: v_max_f64 v[6:7], v[0:1], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_f64__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7f8, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v1, v[0:1]
+; GFX7-NEXT: flat_load_dword v0, v[4:5]
+; GFX7-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v9, v1
+; GFX7-NEXT: v_mov_b32_e32 v8, v0
+; GFX7-NEXT: v_max_f64 v[0:1], v[8:9], v[8:9]
+; GFX7-NEXT: v_max_f64 v[6:7], v[0:1], v[2:3]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr %ptr, i64 255
+ %result = atomicrmw fmax ptr %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define double @flat_agent_atomic_fmax_ret_f64__offset12b_neg(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_f64__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b64 v[4:5], v[0:1] offset:-2048
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[6:7], v[6:7]
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[4:5], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB18_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_f64__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v4
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v5, vcc
+; GFX940-NEXT: flat_load_dwordx2 v[0:1], v[0:1]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[8:9], v[0:1]
+; GFX940-NEXT: v_max_f64 v[0:1], v[8:9], v[8:9]
+; GFX940-NEXT: v_max_f64 v[6:7], v[0:1], v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB18_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_f64__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v5, v0
+; GFX11-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v5
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v4, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v5, vcc_lo, 0xfffff800, v5
+; GFX11-NEXT: v_add_co_ci_u32_e32 v6, vcc_lo, -1, v4, vcc_lo
+; GFX11-NEXT: flat_load_b64 v[0:1], v[0:1]
+; GFX11-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v10, v1 :: v_dual_mov_b32 v9, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[0:1], v[9:10], v[9:10]
+; GFX11-NEXT: v_max_f64 v[7:8], v[0:1], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[0:1], v[5:6], v[7:10] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[9:10]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB18_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_f64__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dwordx2 v[0:1], v[4:5]
+; GFX10-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v9, v1
+; GFX10-NEXT: v_mov_b32_e32 v8, v0
+; GFX10-NEXT: v_max_f64 v[0:1], v[8:9], v[8:9]
+; GFX10-NEXT: v_max_f64 v[6:7], v[0:1], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[8:9]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB18_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_f64__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dwordx2 v[0:1], v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[8:9], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: v_max_f64 v[0:1], v[8:9], v[8:9]
+; GFX90A-NEXT: v_max_f64 v[6:7], v[0:1], v[2:3]
+; GFX90A-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB18_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_f64__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v5, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_mov_b32_e32 v4, v1
+; GFX908-NEXT: s_mov_b64 s[4:5], vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v4, vcc
+; GFX908-NEXT: flat_load_dwordx2 v[0:1], v[0:1]
+; GFX908-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX908-NEXT: v_addc_co_u32_e64 v6, vcc, -1, v4, s[4:5]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v10, v1
+; GFX908-NEXT: v_mov_b32_e32 v9, v0
+; GFX908-NEXT: v_max_f64 v[0:1], v[9:10], v[9:10]
+; GFX908-NEXT: v_max_f64 v[7:8], v[0:1], v[2:3]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[5:6], v[7:10] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[9:10]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB18_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_f64__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff804, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v1, v[0:1]
+; GFX8-NEXT: flat_load_dword v0, v[4:5]
+; GFX8-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v9, v1
+; GFX8-NEXT: v_mov_b32_e32 v8, v0
+; GFX8-NEXT: v_max_f64 v[0:1], v[8:9], v[8:9]
+; GFX8-NEXT: v_max_f64 v[6:7], v[0:1], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB18_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_f64__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff804, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v1, v[0:1]
+; GFX7-NEXT: flat_load_dword v0, v[4:5]
+; GFX7-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v9, v1
+; GFX7-NEXT: v_mov_b32_e32 v8, v0
+; GFX7-NEXT: v_max_f64 v[0:1], v[8:9], v[8:9]
+; GFX7-NEXT: v_max_f64 v[6:7], v[0:1], v[2:3]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB18_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr %ptr, i64 -256
+ %result = atomicrmw fmax ptr %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define void @flat_agent_atomic_fmax_noret_f64(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b64 v[4:5], v[0:1]
+; GFX12-NEXT: v_max_num_f64_e32 v[6:7], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[4:5], v[4:5]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[2:3], v[6:7]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX12-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB19_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX940-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB19_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b64 v[4:5], v[0:1]
+; GFX11-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX11-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB19_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX10-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX10-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX10-NEXT: v_mov_b32_e32 v5, v3
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB19_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX90A-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX90A-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB19_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX908-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX908-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v5, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB19_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 4, v0
+; GFX8-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: flat_load_dword v5, v[5:6]
+; GFX8-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX8-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v5, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB19_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v5, vcc, 4, v0
+; GFX7-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: flat_load_dword v5, v[5:6]
+; GFX7-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX7-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB19_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmax ptr %ptr, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmax_noret_f64__offset12b_pos(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_f64__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b64 v[4:5], v[0:1] offset:2040
+; GFX12-NEXT: v_max_num_f64_e32 v[6:7], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[4:5], v[4:5]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[2:3], v[6:7]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5] offset:2040 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX12-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB20_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_f64__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dwordx2 v[4:5], v[0:1] offset:2040
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX940-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] offset:2040 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB20_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_f64__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b64 v[4:5], v[0:1] offset:2040
+; GFX11-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5] offset:2040 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX11-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB20_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_f64__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7f8, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX10-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX10-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX10-NEXT: v_mov_b32_e32 v5, v3
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB20_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_f64__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dwordx2 v[4:5], v[0:1] offset:2040
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX90A-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX90A-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] offset:2040 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB20_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_f64__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dwordx2 v[4:5], v[0:1] offset:2040
+; GFX908-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX908-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] offset:2040 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v5, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB20_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_f64__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7f8, v0
+; GFX8-NEXT: v_addc_u32_e32 v7, vcc, 0, v1, vcc
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: flat_load_dword v4, v[6:7]
+; GFX8-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX8-NEXT: v_max_f64 v[2:3], v[2:3], v[0:1]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[6:7], v[2:5] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v5, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB20_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_f64__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v6, vcc, 0x7f8, v0
+; GFX7-NEXT: v_addc_u32_e32 v7, vcc, 0, v1, vcc
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: flat_load_dword v4, v[6:7]
+; GFX7-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX7-NEXT: v_max_f64 v[2:3], v[2:3], v[0:1]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[6:7], v[2:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB20_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr %ptr, i64 255
+ %unused = atomicrmw fmax ptr %gep, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmax_noret_f64__offset12b_neg(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_f64__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b64 v[4:5], v[0:1] offset:-2048
+; GFX12-NEXT: v_max_num_f64_e32 v[6:7], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[4:5], v[4:5]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[2:3], v[6:7]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX12-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB21_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_f64__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX940-NEXT: flat_load_dwordx2 v[4:5], v[4:5]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX940-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB21_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_f64__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: flat_load_b64 v[4:5], v[4:5]
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX11-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB21_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_f64__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX10-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX10-NEXT: v_max_f64 v[2:3], v[2:3], v[6:7]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX10-NEXT: v_mov_b32_e32 v5, v3
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB21_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_f64__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v6, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v7, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX90A-NEXT: v_max_f64 v[2:3], v[2:3], v[0:1]
+; GFX90A-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[6:7], v[2:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB21_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_f64__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v6, vcc, 0xfffff800, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], vcc
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dwordx2 v[4:5], v[4:5]
+; GFX908-NEXT: v_max_f64 v[8:9], v[2:3], v[2:3]
+; GFX908-NEXT: v_addc_co_u32_e64 v7, vcc, -1, v1, s[4:5]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
+; GFX908-NEXT: v_max_f64 v[2:3], v[0:1], v[8:9]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[6:7], v[2:5] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v5, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v0
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB21_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_f64__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v7, vcc, -1, v1, vcc
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff804, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: flat_load_dword v4, v[6:7]
+; GFX8-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX8-NEXT: v_max_f64 v[2:3], v[2:3], v[0:1]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[6:7], v[2:5] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v5, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB21_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_f64__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v6, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v7, vcc, -1, v1, vcc
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff804, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: flat_load_dword v4, v[6:7]
+; GFX7-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX7-NEXT: v_max_f64 v[2:3], v[2:3], v[0:1]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[6:7], v[2:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB21_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr %ptr, i64 -256
+ %unused = atomicrmw fmax ptr %gep, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; half
+; --------------------------------------------------------------------
+
+define half @flat_agent_atomic_fmax_ret_f16(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v0
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB22_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB22_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB22_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB22_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB22_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB22_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB22_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v2, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_max_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v2, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB22_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr %ptr, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+define half @flat_agent_atomic_fmax_ret_f16__offset12b_pos(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB23_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB23_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB23_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB23_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB23_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB23_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB23_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v2, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_max_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v2, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB23_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %result = atomicrmw fmax ptr %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+define half @flat_agent_atomic_fmax_ret_f16__offset12b_neg(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB24_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB24_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB24_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB24_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB24_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB24_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB24_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v2, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_max_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v2, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB24_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 -1024
+ %result = atomicrmw fmax ptr %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+ }
+
+define void @flat_agent_atomic_fmax_noret_f16(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v0
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v4, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v3, v3, v3
+; GFX12-NEXT: v_max_num_f16_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB25_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX940-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX940-NEXT: v_max_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB25_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v4, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX11-NEXT: v_max_f16_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB25_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX10-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX10-NEXT: v_max_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB25_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX90A-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX90A-NEXT: v_max_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB25_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX908-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX908-NEXT: v_max_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB25_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX8-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX8-NEXT: v_max_f16_e32 v3, v3, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB25_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: v_lshl_b32_e32 v3, 0xffff, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_not_b32_e32 v6, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v2, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX7-NEXT: v_max_f32_e32 v3, v3, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v2, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB25_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmax ptr %ptr, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmax_noret_f16__offset12b_pos(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v6, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB26_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX940-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB26_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB26_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB26_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX90A-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB26_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX908-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB26_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX8-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB26_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB26_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %unused = atomicrmw fmax ptr %gep, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmax_noret_f16__offset12b_neg(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v6, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB27_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX940-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB27_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB27_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB27_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX90A-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB27_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX908-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB27_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX8-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB27_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB27_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 -1024
+ %unused = atomicrmw fmax ptr %gep, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define half @flat_agent_atomic_fmax_ret_f16__offset12b_pos__align4(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_pos__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v3, v4, v4
+; GFX12-NEXT: v_max_num_f16_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB28_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_pos__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f16_e32 v3, v5, v5
+; GFX940-NEXT: v_max_f16_e32 v3, v3, v2
+; GFX940-NEXT: v_and_or_b32 v4, v5, s2, v3
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB28_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_pos__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v3, v4, v4
+; GFX11-NEXT: v_max_f16_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB28_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_pos__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v1, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_max_f16_e32 v0, v6, v6
+; GFX10-NEXT: v_max_f16_e32 v0, v0, v1
+; GFX10-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX10-NEXT: v_and_or_b32 v5, 0xffff0000, v6, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB28_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_pos__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f16_e32 v3, v5, v5
+; GFX90A-NEXT: v_max_f16_e32 v3, v3, v2
+; GFX90A-NEXT: v_and_or_b32 v4, v5, s6, v3
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB28_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_pos__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f16_e32 v3, v4, v4
+; GFX908-NEXT: v_max_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_and_or_b32 v3, v4, s6, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB28_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_pos__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v1, v2, v2
+; GFX8-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_max_f16_e32 v0, v6, v6
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX8-NEXT: v_max_f16_e32 v0, v0, v1
+; GFX8-NEXT: v_or_b32_e32 v5, v2, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB28_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_f16__offset12b_pos__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB28_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %result = atomicrmw fmax ptr %gep, half %val syncscope("agent") seq_cst, align 4
+ ret half %result
+}
+
+define void @flat_agent_atomic_fmax_noret_f16__offset12b__align4_pos(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b__align4_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX12-NEXT: v_max_num_f16_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f16_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v4
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB29_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b__align4_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v4
+; GFX940-NEXT: v_and_or_b32 v2, v3, s2, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB29_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b__align4_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX11-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v4
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB29_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b__align4_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v4
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB29_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b__align4_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v4
+; GFX90A-NEXT: v_and_or_b32 v2, v3, s6, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB29_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b__align4_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v4
+; GFX908-NEXT: v_and_or_b32 v2, v3, s6, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB29_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b__align4_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX8-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v4
+; GFX8-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB29_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_f16__offset12b__align4_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v2
+; GFX7-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB29_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %unused = atomicrmw fmax ptr %gep, half %val syncscope("agent") seq_cst, align 4
+ ret void
+}
+
+define half @flat_system_atomic_fmax_ret_f16__offset12b_pos(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fmax_ret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB30_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmax_ret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB30_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmax_ret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB30_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmax_ret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB30_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmax_ret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB30_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmax_ret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB30_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmax_ret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB30_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmax_ret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v2, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_max_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v2, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB30_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %result = atomicrmw fmax ptr %gep, half %val seq_cst
+ ret half %result
+}
+
+define void @flat_system_atomic_fmax_noret_f16__offset12b_pos(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fmax_noret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v6, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB31_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmax_noret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX940-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB31_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmax_noret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB31_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmax_noret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB31_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmax_noret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX90A-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB31_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmax_noret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX908-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB31_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmax_noret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX8-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v6
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB31_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmax_noret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB31_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %unused = atomicrmw fmax ptr %gep, half %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; bfloat
+; --------------------------------------------------------------------
+
+define bfloat @flat_agent_atomic_fmax_ret_bf16(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB32_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB32_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB32_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB32_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB32_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB32_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB32_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB32_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr %ptr, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+define bfloat @flat_agent_atomic_fmax_ret_bf16__offset12b_pos(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB33_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB33_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB33_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB33_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB33_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB33_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB33_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB33_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %result = atomicrmw fmax ptr %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+define bfloat @flat_agent_atomic_fmax_ret_bf16__offset12b_neg(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB34_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB34_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB34_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB34_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB34_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB34_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB34_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB34_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 -1024
+ %result = atomicrmw fmax ptr %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+ }
+
+define void @flat_agent_atomic_fmax_noret_bf16(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v4, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB35_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v4, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_max_f32_e32 v4, v4, v2
+; GFX940-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX940-NEXT: v_add3_u32 v7, v7, v4, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v4, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB35_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v4, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB35_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB35_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_max_f32_e32 v4, v4, v2
+; GFX90A-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX90A-NEXT: v_add3_u32 v7, v7, v4, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB35_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB35_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX8-NEXT: v_bfe_u32 v8, v3, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v3
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB35_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v3, 0xffff, v5
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v6, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB35_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmax ptr %ptr, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmax_noret_bf16__offset12b_pos(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB36_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB36_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB36_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB36_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB36_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB36_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB36_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB36_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %unused = atomicrmw fmax ptr %gep, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmax_noret_bf16__offset12b_neg(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB37_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB37_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB37_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB37_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB37_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB37_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB37_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB37_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 -1024
+ %unused = atomicrmw fmax ptr %gep, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define bfloat @flat_agent_atomic_fmax_ret_bf16__offset12b_pos__align4(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_pos__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB38_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_pos__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX940-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX940-NEXT: v_bfe_u32 v4, v3, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX940-NEXT: v_add3_u32 v4, v4, v3, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v3, v4, v6, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX940-NEXT: v_and_or_b32 v4, v5, s3, v3
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB38_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_pos__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB38_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_pos__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX10-NEXT: v_max_f32_e32 v0, v0, v1
+; GFX10-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX10-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX10-NEXT: v_and_or_b32 v5, 0xffff0000, v6, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB38_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_pos__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX90A-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX90A-NEXT: v_bfe_u32 v4, v3, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX90A-NEXT: v_add3_u32 v4, v4, v3, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e32 v3, v4, v6, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX90A-NEXT: v_and_or_b32 v4, v5, s7, v3
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB38_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_pos__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX908-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, s7, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB38_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_pos__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_max_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v0
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX8-NEXT: v_cndmask_b32_e32 v0, v5, v7, vcc
+; GFX8-NEXT: v_or_b32_sdwa v5, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB38_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_bf16__offset12b_pos__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB38_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %result = atomicrmw fmax ptr %gep, bfloat %val syncscope("agent") seq_cst, align 4
+ ret bfloat %result
+}
+
+define void @flat_agent_atomic_fmax_noret_bf16__offset12b__align4_pos(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b__align4_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB39_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b__align4_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v5, v5, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: v_and_or_b32 v2, v3, s3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB39_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b__align4_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB39_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b__align4_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB39_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b__align4_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v5, v5, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v3, s7, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB39_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b__align4_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v5, v5, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: v_and_or_b32 v2, v3, s7, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB39_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b__align4_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v2
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v6, v7, vcc
+; GFX8-NEXT: v_or_b32_sdwa v2, v5, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB39_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b__align4_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX7-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB39_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %unused = atomicrmw fmax ptr %gep, bfloat %val syncscope("agent") seq_cst, align 4
+ ret void
+}
+
+define bfloat @flat_system_atomic_fmax_ret_bf16__offset12b_pos(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB40_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB40_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB40_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB40_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB40_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB40_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB40_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmax_ret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB40_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %result = atomicrmw fmax ptr %gep, bfloat %val seq_cst
+ ret bfloat %result
+}
+
+define void @flat_system_atomic_fmax_noret_bf16__offset12b_pos(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB41_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB41_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB41_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB41_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB41_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB41_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB41_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmax_noret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_max_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB41_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %unused = atomicrmw fmax ptr %gep, bfloat %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x half>
+; --------------------------------------------------------------------
+
+define <2 x half> @flat_agent_atomic_fmax_ret_v2f16(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v3, v4, v4
+; GFX12-NEXT: v_pk_max_num_f16 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB42_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX940-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_max_f16 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB42_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX11-NEXT: v_pk_max_f16 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB42_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX10-NEXT: v_pk_max_f16 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB42_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX90A-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX90A-NEXT: v_pk_max_f16 v4, v3, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB42_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX908-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX908-NEXT: v_pk_max_f16 v3, v3, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB42_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v4, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v3
+; GFX8-NEXT: v_max_f16_sdwa v3, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v6, v6
+; GFX8-NEXT: v_max_f16_sdwa v3, v3, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB42_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX7-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[6:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB42_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr %ptr, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define <2 x half> @flat_agent_atomic_fmax_ret_v2f16__offset12b_pos(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v3, v4, v4
+; GFX12-NEXT: v_pk_max_num_f16 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB43_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX940-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_max_f16 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB43_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX11-NEXT: v_pk_max_f16 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB43_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_pk_max_f16 v1, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_pk_max_f16 v0, v6, v6
+; GFX10-NEXT: v_pk_max_f16 v5, v0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB43_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX90A-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX90A-NEXT: v_pk_max_f16 v4, v3, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB43_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX908-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX908-NEXT: v_pk_max_f16 v3, v3, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB43_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v1, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_max_f16_sdwa v0, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v6, v6
+; GFX8-NEXT: v_max_f16_sdwa v0, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB43_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v1, v[4:5]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v0
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[4:5], v[6:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB43_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 511
+ %result = atomicrmw fmax ptr %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define <2 x half> @flat_agent_atomic_fmax_ret_v2f16__offset12b_neg(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_v2f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v3, v4, v4
+; GFX12-NEXT: v_pk_max_num_f16 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB44_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_v2f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v4
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v5, vcc
+; GFX940-NEXT: flat_load_dword v0, v[0:1]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v1, v2, v2
+; GFX940-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_max_f16 v0, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_max_f16 v2, v0, v1
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v0, v[4:5], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB44_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_v2f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: flat_load_b32 v0, v[4:5]
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_pk_max_f16 v1, v2, v2
+; GFX11-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v0, v6, v6
+; GFX11-NEXT: v_pk_max_f16 v5, v0, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v0, v[3:4], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB44_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_v2f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_pk_max_f16 v1, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_pk_max_f16 v0, v6, v6
+; GFX10-NEXT: v_pk_max_f16 v5, v0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB44_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_v2f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v0, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v1, v2, v2
+; GFX90A-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_max_f16 v0, v3, v3
+; GFX90A-NEXT: v_pk_max_f16 v2, v0, v1
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB44_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_v2f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v0, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v1, v2, v2
+; GFX908-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v0
+; GFX908-NEXT: v_pk_max_f16 v0, v6, v6
+; GFX908-NEXT: v_pk_max_f16 v5, v0, v1
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB44_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_v2f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v1, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_max_f16_sdwa v0, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v6, v6
+; GFX8-NEXT: v_max_f16_sdwa v0, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB44_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_v2f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v1, v[4:5]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v0
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[4:5], v[6:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB44_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 -512
+ %result = atomicrmw fmax ptr %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define void @flat_agent_atomic_fmax_noret_v2f16(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_pk_max_num_f16 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB45_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX940-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB45_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB45_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX10-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB45_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX90A-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB45_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX908-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB45_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v4, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v2, v2
+; GFX8-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v2, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v6, v3, v3
+; GFX8-NEXT: v_max_f16_sdwa v2, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v6, v6, v5
+; GFX8-NEXT: v_or_b32_e32 v2, v6, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB45_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: flat_atomic_cmpswap v7, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB45_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmax ptr %ptr, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmax_noret_v2f16__offset12b_pos(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_pk_max_num_f16 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB46_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX940-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB46_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB46_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX10-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB46_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX90A-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB46_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX908-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB46_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v4, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v2, v2
+; GFX8-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v2, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v6, v3, v3
+; GFX8-NEXT: v_max_f16_sdwa v2, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v6, v6, v5
+; GFX8-NEXT: v_or_b32_e32 v2, v6, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB46_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: flat_atomic_cmpswap v7, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB46_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 511
+ %unused = atomicrmw fmax ptr %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmax_noret_v2f16__offset12b_neg(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_v2f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: v_pk_max_num_f16 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB47_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_v2f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX940-NEXT: flat_load_dword v3, v[4:5]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX940-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB47_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_v2f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: flat_load_b32 v3, v[3:4]
+; GFX11-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB47_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_v2f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX10-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB47_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_v2f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v1, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX90A-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v0, v1, v1
+; GFX90A-NEXT: v_pk_max_f16 v0, v0, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB47_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_v2f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v1, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX908-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v0, v1, v1
+; GFX908-NEXT: v_pk_max_f16 v0, v0, v2
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB47_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_v2f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v4, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v2, v2
+; GFX8-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v2, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v6, v3, v3
+; GFX8-NEXT: v_max_f16_sdwa v2, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v6, v6, v5
+; GFX8-NEXT: v_or_b32_e32 v2, v6, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB47_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_v2f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: flat_atomic_cmpswap v7, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB47_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 -512
+ %unused = atomicrmw fmax ptr %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x half> @flat_system_atomic_fmax_ret_v2f16__offset12b_pos(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v3, v4, v4
+; GFX12-NEXT: v_pk_max_num_f16 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB48_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX940-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_max_f16 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB48_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX11-NEXT: v_pk_max_f16 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB48_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_pk_max_f16 v1, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_pk_max_f16 v0, v6, v6
+; GFX10-NEXT: v_pk_max_f16 v5, v0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB48_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX90A-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX90A-NEXT: v_pk_max_f16 v4, v3, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB48_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX908-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX908-NEXT: v_pk_max_f16 v3, v3, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB48_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v1, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_max_f16_sdwa v0, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v6, v6
+; GFX8-NEXT: v_max_f16_sdwa v0, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB48_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmax_ret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v1, v[4:5]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v0
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[4:5], v[6:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB48_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 511
+ %result = atomicrmw fmax ptr %gep, <2 x half> %val seq_cst
+ ret <2 x half> %result
+}
+
+define void @flat_system_atomic_fmax_noret_v2f16__offset12b_pos(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_pk_max_num_f16 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB49_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX940-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB49_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB49_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX10-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB49_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX90A-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB49_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX908-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v4
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB49_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v4, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v2, v2
+; GFX8-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v2, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v6, v3, v3
+; GFX8-NEXT: v_max_f16_sdwa v2, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v6, v6, v5
+; GFX8-NEXT: v_or_b32_e32 v2, v6, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB49_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmax_noret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: flat_atomic_cmpswap v7, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB49_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 511
+ %unused = atomicrmw fmax ptr %gep, <2 x half> %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x bfloat>
+; --------------------------------------------------------------------
+
+define <2 x bfloat> @flat_agent_atomic_fmax_ret_v2bf16(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_max_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB50_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB50_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB50_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB50_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB50_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB50_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v3
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v3, 16
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB50_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v2, v2, v7, 16
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB50_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmax ptr %ptr, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define <2 x bfloat> @flat_agent_atomic_fmax_ret_v2bf16__offset12b_pos(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_max_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB51_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB51_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB51_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_max_f32_e32 v0, v0, v1
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v0
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v0, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v0, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB51_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB51_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB51_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_max_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB51_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[4:5]
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v0, v0, v7, 16
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB51_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 511
+ %result = atomicrmw fmax ptr %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define <2 x bfloat> @flat_agent_atomic_fmax_ret_v2bf16__offset12b_neg(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_ret_v2bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_max_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB52_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_ret_v2bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v4
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v5, vcc
+; GFX940-NEXT: flat_load_dword v0, v[0:1]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v3, 0xffff0000, v7
+; GFX940-NEXT: v_max_f32_e32 v0, v0, v1
+; GFX940-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX940-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v3, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v3
+; GFX940-NEXT: v_add3_u32 v6, v6, v0, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v3, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v0, v0
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v0, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v3, v0, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v0, v[4:5], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB52_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_ret_v2bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: flat_load_b32 v0, v[4:5]
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_max_f32_e32 v0, v0, v1
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v0
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v0, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v0, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v0, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v0, v[3:4], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB52_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_ret_v2bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_max_f32_e32 v0, v0, v1
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v0
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v0, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v0, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB52_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_ret_v2bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v0, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v3, 0xffff0000, v7
+; GFX90A-NEXT: v_max_f32_e32 v0, v0, v1
+; GFX90A-NEXT: v_max_f32_e32 v3, v3, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v3, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v3
+; GFX90A-NEXT: v_add3_u32 v6, v6, v0, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v3, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX90A-NEXT: v_cndmask_b32_e64 v0, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v3, v0, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB52_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_ret_v2bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v0, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_max_f32_e32 v0, v0, v1
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v0, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX908-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v0, s9
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB52_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_ret_v2bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_max_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB52_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_ret_v2bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[4:5]
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v0, v0, v7, 16
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB52_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 -512
+ %result = atomicrmw fmax ptr %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @flat_agent_atomic_fmax_noret_v2bf16(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_max_num_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB53_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB53_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB53_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB53_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB53_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB53_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB53_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX7-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB53_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmax ptr %ptr, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmax_noret_v2bf16__offset12b_pos(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_max_num_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB54_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB54_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB54_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB54_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB54_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB54_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB54_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX7-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB54_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 511
+ %unused = atomicrmw fmax ptr %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmax_noret_v2bf16__offset12b_neg(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmax_noret_v2bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_max_num_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB55_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmax_noret_v2bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX940-NEXT: flat_load_dword v3, v[4:5]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB55_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmax_noret_v2bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: flat_load_b32 v3, v[3:4]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB55_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmax_noret_v2bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB55_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmax_noret_v2bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v1, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX90A-NEXT: v_max_f32_e32 v0, v0, v3
+; GFX90A-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX90A-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v0, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX90A-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v0, v6, v0, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB55_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmax_noret_v2bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v1, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX908-NEXT: v_max_f32_e32 v0, v0, v5
+; GFX908-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX908-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v0, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX908-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v0, v6, v0, s9
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB55_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmax_noret_v2bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB55_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmax_noret_v2bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX7-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB55_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 -512
+ %unused = atomicrmw fmax ptr %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x bfloat> @flat_system_atomic_fmax_ret_v2bf16__offset12b_pos(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_max_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_max_num_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB56_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB56_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB56_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_max_f32_e32 v0, v0, v1
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v0
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v0, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v0, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB56_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB56_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_max_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB56_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_max_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB56_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmax_ret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[4:5]
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v0, v0, v7, 16
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB56_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 511
+ %result = atomicrmw fmax ptr %gep, <2 x bfloat> %val seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @flat_system_atomic_fmax_noret_v2bf16__offset12b_pos(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_max_num_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB57_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB57_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB57_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB57_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB57_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB57_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_max_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_max_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB57_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmax_noret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX7-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB57_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 511
+ %unused = atomicrmw fmax ptr %gep, <2 x bfloat> %val seq_cst
+ ret void
+}
+
+attributes #0 = { nounwind "amdgpu-unsafe-fp-atomics"="true" }
+attributes #1 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" "amdgpu-unsafe-fp-atomics"="true" }
diff --git a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmin.ll b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmin.ll
new file mode 100644
index 0000000000000..fdfbb4215eb6f
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmin.ll
@@ -0,0 +1,15756 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 < %s | FileCheck -check-prefix=GFX12 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx940 < %s | FileCheck -check-prefix=GFX940 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx90a < %s | FileCheck -check-prefix=GFX90A %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx908 < %s | FileCheck -check-prefix=GFX908 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tonga < %s | FileCheck -check-prefix=GFX8 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=hawaii < %s | FileCheck -check-prefix=GFX7 %s
+
+; --------------------------------------------------------------------
+; float
+; --------------------------------------------------------------------
+
+define float @flat_agent_atomic_fmin_ret_f32(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB0_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB0_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB0_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX10-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB0_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB0_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB0_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX8-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_mul_f32_e32 v3, 1.0, v4
+; GFX8-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB0_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v4
+; GFX7-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB0_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr %ptr, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @flat_agent_atomic_fmin_ret_f32__offset12b_pos(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB1_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB1_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB1_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_max_f32_e32 v0, v6, v6
+; GFX10-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB1_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB1_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB1_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB1_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX7-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB1_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %result = atomicrmw fmin ptr %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @flat_agent_atomic_fmin_ret_f32__offset12b_neg(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB2_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v4
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v5, vcc
+; GFX940-NEXT: flat_load_dword v0, v[0:1]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX940-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_max_f32_e32 v0, v3, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v0, v1
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v0, v[4:5], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB2_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: flat_load_b32 v0, v[4:5]
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX11-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v0, v6, v6
+; GFX11-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v0, v[3:4], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB2_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_max_f32_e32 v0, v6, v6
+; GFX10-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB2_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v0, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX90A-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_max_f32_e32 v0, v3, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v0, v1
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB2_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v0, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX908-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v0
+; GFX908-NEXT: v_max_f32_e32 v0, v6, v6
+; GFX908-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB2_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB2_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX7-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 -512
+ %result = atomicrmw fmin ptr %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define void @flat_agent_atomic_fmin_noret_f32(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB3_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB3_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB3_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB3_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB3_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB3_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB3_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB3_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmin ptr %ptr, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmin_noret_f32__offset12b_pos(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB4_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB4_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB4_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB4_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB4_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB4_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB4_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB4_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %unused = atomicrmw fmin ptr %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmin_noret_f32__offset12b_neg(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB5_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX940-NEXT: flat_load_dword v3, v[4:5]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB5_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: flat_load_b32 v3, v[3:4]
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB5_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB5_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v1, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v0, v1, v1
+; GFX90A-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB5_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v1, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v0, v1, v1
+; GFX908-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 -512
+ %unused = atomicrmw fmin ptr %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define float @flat_system_atomic_fmin_ret_f32__offset12b_pos(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fmin_ret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB6_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmin_ret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB6_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmin_ret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB6_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmin_ret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_max_f32_e32 v0, v6, v6
+; GFX10-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB6_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmin_ret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB6_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmin_ret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB6_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmin_ret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB6_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmin_ret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX7-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB6_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %result = atomicrmw fmin ptr %gep, float %val seq_cst
+ ret float %result
+}
+
+define void @flat_system_atomic_fmin_noret_f32__offset12b_pos(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fmin_noret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB7_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmin_noret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB7_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmin_noret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB7_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmin_noret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB7_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmin_noret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB7_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmin_noret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB7_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmin_noret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB7_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmin_noret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB7_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %unused = atomicrmw fmin ptr %gep, float %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; float with ftz/daz
+; --------------------------------------------------------------------
+
+define float @flat_agent_atomic_fmin_ret_f32__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_f32__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB8_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_f32__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB8_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_f32__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB8_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_f32__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX10-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB8_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_f32__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB8_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_f32__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_f32__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX8-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_mul_f32_e32 v3, 1.0, v4
+; GFX8-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_f32__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v4
+; GFX7-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB8_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr %ptr, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @flat_agent_atomic_fmin_ret_f32__offset12b_pos__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB9_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB9_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB9_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_max_f32_e32 v0, v6, v6
+; GFX10-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB9_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB9_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB9_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB9_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX7-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB9_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %result = atomicrmw fmin ptr %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @flat_agent_atomic_fmin_ret_f32__offset12b_neg__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_neg__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB10_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_neg__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v4
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v5, vcc
+; GFX940-NEXT: flat_load_dword v0, v[0:1]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX940-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_max_f32_e32 v0, v3, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v0, v1
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v0, v[4:5], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB10_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_neg__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: flat_load_b32 v0, v[4:5]
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX11-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v0, v6, v6
+; GFX11-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v0, v[3:4], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB10_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_neg__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_max_f32_e32 v0, v6, v6
+; GFX10-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB10_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_neg__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v0, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX90A-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_max_f32_e32 v0, v3, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v0, v1
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB10_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_neg__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v0, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX908-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v0
+; GFX908-NEXT: v_max_f32_e32 v0, v6, v6
+; GFX908-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB10_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_neg__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB10_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_f32__offset12b_neg__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX7-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB10_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 -512
+ %result = atomicrmw fmin ptr %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define void @flat_agent_atomic_fmin_noret_f32__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_f32__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB11_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_f32__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB11_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_f32__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB11_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_f32__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB11_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_f32__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_f32__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_f32__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_f32__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB11_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmin ptr %ptr, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmin_noret_f32__offset12b_pos__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB12_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB12_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB12_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB12_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB12_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB12_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB12_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB12_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %unused = atomicrmw fmin ptr %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmin_noret_f32__offset12b_neg__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_neg__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB13_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_neg__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX940-NEXT: flat_load_dword v3, v[4:5]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB13_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_neg__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: flat_load_b32 v3, v[3:4]
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB13_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_neg__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB13_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_neg__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v1, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v0, v1, v1
+; GFX90A-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB13_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_neg__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v1, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v0, v1, v1
+; GFX908-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB13_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_neg__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB13_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_f32__offset12b_neg__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB13_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 -512
+ %unused = atomicrmw fmin ptr %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define float @flat_system_atomic_fmin_ret_f32__offset12b_pos__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_system_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v3, v4, v4
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB14_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX940-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX940-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB14_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB14_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v1, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_max_f32_e32 v0, v6, v6
+; GFX10-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB14_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f32_e32 v3, v5, v5
+; GFX90A-NEXT: v_min_f32_e32 v4, v3, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v2, v2
+; GFX908-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f32_e32 v3, v4, v4
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX8-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX8-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmin_ret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v6
+; GFX7-NEXT: v_min_f32_e32 v5, v0, v1
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %result = atomicrmw fmin ptr %gep, float %val seq_cst
+ ret float %result
+}
+
+define void @flat_system_atomic_fmin_noret_f32__offset12b_pos__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_system_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_max_num_f32_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB15_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX940-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB15_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB15_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB15_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX90A-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB15_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f32_e32 v4, v2, v2
+; GFX908-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v2, v3, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB15_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX8-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB15_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmin_noret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v2
+; GFX7-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB15_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %unused = atomicrmw fmin ptr %gep, float %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; double
+; --------------------------------------------------------------------
+
+define double @flat_agent_atomic_fmin_ret_f64(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b64 v[4:5], v[0:1]
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[6:7], v[6:7]
+; GFX12-NEXT: v_min_num_f64_e32 v[4:5], v[4:5], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB16_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX940-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB16_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b64 v[4:5], v[0:1]
+; GFX11-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX11-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB16_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX10-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX10-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB16_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX90A-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX90A-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB16_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX908-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX908-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB16_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 4, v0
+; GFX8-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: flat_load_dword v5, v[5:6]
+; GFX8-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v7, v5
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX8-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB16_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: v_mov_b32_e32 v1, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v5, vcc, 4, v0
+; GFX7-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: flat_load_dword v5, v[5:6]
+; GFX7-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX7-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB16_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v4
+; GFX7-NEXT: v_mov_b32_e32 v1, v5
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr %ptr, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define double @flat_agent_atomic_fmin_ret_f64__offset12b_pos(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_f64__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b64 v[4:5], v[0:1] offset:2040
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[6:7], v[6:7]
+; GFX12-NEXT: v_min_num_f64_e32 v[4:5], v[4:5], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] offset:2040 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB17_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_f64__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dwordx2 v[4:5], v[0:1] offset:2040
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX940-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] offset:2040 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB17_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_f64__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b64 v[4:5], v[0:1] offset:2040
+; GFX11-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX11-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] offset:2040 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB17_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_f64__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7f8, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dwordx2 v[0:1], v[4:5]
+; GFX10-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v9, v1
+; GFX10-NEXT: v_mov_b32_e32 v8, v0
+; GFX10-NEXT: v_max_f64 v[0:1], v[8:9], v[8:9]
+; GFX10-NEXT: v_min_f64 v[6:7], v[0:1], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[8:9]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB17_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_f64__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dwordx2 v[4:5], v[0:1] offset:2040
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX90A-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX90A-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] offset:2040 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB17_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_f64__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dwordx2 v[4:5], v[0:1] offset:2040
+; GFX908-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: v_max_f64 v[4:5], v[6:7], v[6:7]
+; GFX908-NEXT: v_min_f64 v[4:5], v[4:5], v[2:3]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] offset:2040 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_f64__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7f8, v0
+; GFX8-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v1, v[0:1]
+; GFX8-NEXT: flat_load_dword v0, v[4:5]
+; GFX8-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v9, v1
+; GFX8-NEXT: v_mov_b32_e32 v8, v0
+; GFX8-NEXT: v_max_f64 v[0:1], v[8:9], v[8:9]
+; GFX8-NEXT: v_min_f64 v[6:7], v[0:1], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_f64__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7f8, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v1, v[0:1]
+; GFX7-NEXT: flat_load_dword v0, v[4:5]
+; GFX7-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v9, v1
+; GFX7-NEXT: v_mov_b32_e32 v8, v0
+; GFX7-NEXT: v_max_f64 v[0:1], v[8:9], v[8:9]
+; GFX7-NEXT: v_min_f64 v[6:7], v[0:1], v[2:3]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr %ptr, i64 255
+ %result = atomicrmw fmin ptr %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define double @flat_agent_atomic_fmin_ret_f64__offset12b_neg(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_f64__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b64 v[4:5], v[0:1] offset:-2048
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[6:7], v[6:7]
+; GFX12-NEXT: v_min_num_f64_e32 v[4:5], v[4:5], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB18_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_f64__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v4
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v5, vcc
+; GFX940-NEXT: flat_load_dwordx2 v[0:1], v[0:1]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[8:9], v[0:1]
+; GFX940-NEXT: v_max_f64 v[0:1], v[8:9], v[8:9]
+; GFX940-NEXT: v_min_f64 v[6:7], v[0:1], v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB18_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_f64__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v5, v0
+; GFX11-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v5
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v4, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v5, vcc_lo, 0xfffff800, v5
+; GFX11-NEXT: v_add_co_ci_u32_e32 v6, vcc_lo, -1, v4, vcc_lo
+; GFX11-NEXT: flat_load_b64 v[0:1], v[0:1]
+; GFX11-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v10, v1 :: v_dual_mov_b32 v9, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[0:1], v[9:10], v[9:10]
+; GFX11-NEXT: v_min_f64 v[7:8], v[0:1], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[0:1], v[5:6], v[7:10] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[9:10]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB18_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_f64__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dwordx2 v[0:1], v[4:5]
+; GFX10-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v9, v1
+; GFX10-NEXT: v_mov_b32_e32 v8, v0
+; GFX10-NEXT: v_max_f64 v[0:1], v[8:9], v[8:9]
+; GFX10-NEXT: v_min_f64 v[6:7], v[0:1], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[8:9]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB18_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_f64__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dwordx2 v[0:1], v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[8:9], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: v_max_f64 v[0:1], v[8:9], v[8:9]
+; GFX90A-NEXT: v_min_f64 v[6:7], v[0:1], v[2:3]
+; GFX90A-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB18_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_f64__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v5, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_mov_b32_e32 v4, v1
+; GFX908-NEXT: s_mov_b64 s[4:5], vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v4, vcc
+; GFX908-NEXT: flat_load_dwordx2 v[0:1], v[0:1]
+; GFX908-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX908-NEXT: v_addc_co_u32_e64 v6, vcc, -1, v4, s[4:5]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v10, v1
+; GFX908-NEXT: v_mov_b32_e32 v9, v0
+; GFX908-NEXT: v_max_f64 v[0:1], v[9:10], v[9:10]
+; GFX908-NEXT: v_min_f64 v[7:8], v[0:1], v[2:3]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[5:6], v[7:10] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[9:10]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB18_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_f64__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff804, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v1, v[0:1]
+; GFX8-NEXT: flat_load_dword v0, v[4:5]
+; GFX8-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v9, v1
+; GFX8-NEXT: v_mov_b32_e32 v8, v0
+; GFX8-NEXT: v_max_f64 v[0:1], v[8:9], v[8:9]
+; GFX8-NEXT: v_min_f64 v[6:7], v[0:1], v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB18_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_f64__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff804, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v1, v[0:1]
+; GFX7-NEXT: flat_load_dword v0, v[4:5]
+; GFX7-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v9, v1
+; GFX7-NEXT: v_mov_b32_e32 v8, v0
+; GFX7-NEXT: v_max_f64 v[0:1], v[8:9], v[8:9]
+; GFX7-NEXT: v_min_f64 v[6:7], v[0:1], v[2:3]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB18_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr %ptr, i64 -256
+ %result = atomicrmw fmin ptr %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define void @flat_agent_atomic_fmin_noret_f64(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b64 v[4:5], v[0:1]
+; GFX12-NEXT: v_max_num_f64_e32 v[6:7], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[4:5], v[4:5]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f64_e32 v[2:3], v[2:3], v[6:7]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX12-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB19_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX940-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB19_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b64 v[4:5], v[0:1]
+; GFX11-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX11-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB19_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX10-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX10-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX10-NEXT: v_mov_b32_e32 v5, v3
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB19_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX90A-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX90A-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB19_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX908-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX908-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v5, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB19_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 4, v0
+; GFX8-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: flat_load_dword v5, v[5:6]
+; GFX8-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX8-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v5, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB19_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v5, vcc, 4, v0
+; GFX7-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: flat_load_dword v5, v[5:6]
+; GFX7-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX7-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB19_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmin ptr %ptr, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmin_noret_f64__offset12b_pos(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_f64__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b64 v[4:5], v[0:1] offset:2040
+; GFX12-NEXT: v_max_num_f64_e32 v[6:7], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[4:5], v[4:5]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f64_e32 v[2:3], v[2:3], v[6:7]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5] offset:2040 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX12-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB20_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_f64__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dwordx2 v[4:5], v[0:1] offset:2040
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX940-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] offset:2040 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB20_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_f64__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b64 v[4:5], v[0:1] offset:2040
+; GFX11-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5] offset:2040 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX11-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB20_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_f64__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7f8, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX10-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX10-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX10-NEXT: v_mov_b32_e32 v5, v3
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB20_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_f64__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dwordx2 v[4:5], v[0:1] offset:2040
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX90A-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX90A-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] offset:2040 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB20_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_f64__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dwordx2 v[4:5], v[0:1] offset:2040
+; GFX908-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX908-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] offset:2040 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v5, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB20_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_f64__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7f8, v0
+; GFX8-NEXT: v_addc_u32_e32 v7, vcc, 0, v1, vcc
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: flat_load_dword v4, v[6:7]
+; GFX8-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX8-NEXT: v_min_f64 v[2:3], v[2:3], v[0:1]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[6:7], v[2:5] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v5, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB20_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_f64__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v6, vcc, 0x7f8, v0
+; GFX7-NEXT: v_addc_u32_e32 v7, vcc, 0, v1, vcc
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: flat_load_dword v4, v[6:7]
+; GFX7-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX7-NEXT: v_min_f64 v[2:3], v[2:3], v[0:1]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[6:7], v[2:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB20_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr %ptr, i64 255
+ %unused = atomicrmw fmin ptr %gep, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmin_noret_f64__offset12b_neg(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_f64__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b64 v[4:5], v[0:1] offset:-2048
+; GFX12-NEXT: v_max_num_f64_e32 v[6:7], v[2:3], v[2:3]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f64_e32 v[2:3], v[4:5], v[4:5]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f64_e32 v[2:3], v[2:3], v[6:7]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX12-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB21_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_f64__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX940-NEXT: flat_load_dwordx2 v[4:5], v[4:5]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX940-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX940-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB21_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_f64__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: flat_load_b64 v[4:5], v[4:5]
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[2:3], v[0:1], v[2:5] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX11-NEXT: v_dual_mov_b32 v5, v3 :: v_dual_mov_b32 v4, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB21_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_f64__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_max_f64 v[6:7], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX10-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX10-NEXT: v_min_f64 v[2:3], v[2:3], v[6:7]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[0:1], v[2:5] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[4:5]
+; GFX10-NEXT: v_mov_b32_e32 v5, v3
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB21_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_f64__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v6, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v7, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX90A-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX90A-NEXT: v_min_f64 v[2:3], v[2:3], v[0:1]
+; GFX90A-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[6:7], v[2:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB21_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_f64__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v6, vcc, 0xfffff800, v0
+; GFX908-NEXT: s_mov_b64 s[4:5], vcc
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dwordx2 v[4:5], v[4:5]
+; GFX908-NEXT: v_max_f64 v[8:9], v[2:3], v[2:3]
+; GFX908-NEXT: v_addc_co_u32_e64 v7, vcc, -1, v1, s[4:5]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
+; GFX908-NEXT: v_min_f64 v[2:3], v[0:1], v[8:9]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[6:7], v[2:5] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v5, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v0
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB21_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_f64__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v7, vcc, -1, v1, vcc
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff804, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: flat_load_dword v4, v[6:7]
+; GFX8-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX8-NEXT: v_min_f64 v[2:3], v[2:3], v[0:1]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[6:7], v[2:5] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v5, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB21_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_f64__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v6, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v7, vcc, -1, v1, vcc
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff804, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: flat_load_dword v4, v[6:7]
+; GFX7-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
+; GFX7-NEXT: v_min_f64 v[2:3], v[2:3], v[0:1]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[2:3], v[6:7], v[2:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB21_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr %ptr, i64 -256
+ %unused = atomicrmw fmin ptr %gep, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; half
+; --------------------------------------------------------------------
+
+define half @flat_agent_atomic_fmin_ret_f16(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v0
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f16_e32 v5, v5, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB22_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX940-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB22_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB22_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX10-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB22_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX90A-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB22_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX908-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB22_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB22_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v2, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_min_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v2, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB22_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr %ptr, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+define half @flat_agent_atomic_fmin_ret_f16__offset12b_pos(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f16_e32 v5, v5, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB23_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX940-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB23_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB23_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX10-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB23_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX90A-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB23_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX908-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB23_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB23_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v2, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_min_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v2, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB23_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %result = atomicrmw fmin ptr %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+define half @flat_agent_atomic_fmin_ret_f16__offset12b_neg(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f16_e32 v5, v5, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB24_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX940-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB24_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB24_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX10-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB24_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX90A-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB24_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX908-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB24_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB24_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v2, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_min_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v2, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB24_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 -1024
+ %result = atomicrmw fmin ptr %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+ }
+
+define void @flat_agent_atomic_fmin_noret_f16(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v0
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v4, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v3, v3, v3
+; GFX12-NEXT: v_min_num_f16_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB25_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX940-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX940-NEXT: v_min_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB25_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v4, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX11-NEXT: v_min_f16_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB25_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX10-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX10-NEXT: v_min_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB25_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX90A-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX90A-NEXT: v_min_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB25_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX908-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX908-NEXT: v_min_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB25_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX8-NEXT: v_max_f16_e32 v3, v3, v3
+; GFX8-NEXT: v_min_f16_e32 v3, v3, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB25_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: v_lshl_b32_e32 v3, 0xffff, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_not_b32_e32 v6, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v2, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX7-NEXT: v_min_f32_e32 v3, v3, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v2, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB25_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmin ptr %ptr, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmin_noret_f16__offset12b_pos(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v6, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: v_min_num_f16_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB26_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX940-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB26_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB26_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB26_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX90A-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB26_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX908-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB26_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX8-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB26_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB26_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %unused = atomicrmw fmin ptr %gep, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmin_noret_f16__offset12b_neg(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v6, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: v_min_num_f16_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB27_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX940-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB27_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB27_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB27_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX90A-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB27_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX908-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB27_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX8-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB27_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB27_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 -1024
+ %unused = atomicrmw fmin ptr %gep, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define half @flat_agent_atomic_fmin_ret_f16__offset12b_pos__align4(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_pos__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v3, v4, v4
+; GFX12-NEXT: v_min_num_f16_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB28_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_pos__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_max_f16_e32 v3, v5, v5
+; GFX940-NEXT: v_min_f16_e32 v3, v3, v2
+; GFX940-NEXT: v_and_or_b32 v4, v5, s2, v3
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB28_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_pos__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v3, v4, v4
+; GFX11-NEXT: v_min_f16_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB28_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_pos__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v1, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_max_f16_e32 v0, v6, v6
+; GFX10-NEXT: v_min_f16_e32 v0, v0, v1
+; GFX10-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX10-NEXT: v_and_or_b32 v5, 0xffff0000, v6, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB28_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_pos__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_max_f16_e32 v3, v5, v5
+; GFX90A-NEXT: v_min_f16_e32 v3, v3, v2
+; GFX90A-NEXT: v_and_or_b32 v4, v5, s6, v3
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB28_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_pos__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_max_f16_e32 v3, v4, v4
+; GFX908-NEXT: v_min_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_and_or_b32 v3, v4, s6, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB28_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_pos__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v1, v2, v2
+; GFX8-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_max_f16_e32 v0, v6, v6
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX8-NEXT: v_min_f16_e32 v0, v0, v1
+; GFX8-NEXT: v_or_b32_e32 v5, v2, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB28_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_f16__offset12b_pos__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB28_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %result = atomicrmw fmin ptr %gep, half %val syncscope("agent") seq_cst, align 4
+ ret half %result
+}
+
+define void @flat_agent_atomic_fmin_noret_f16__offset12b__align4_pos(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b__align4_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX12-NEXT: v_max_num_f16_e32 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_max_num_f16_e32 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f16_e32 v2, v2, v4
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB29_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b__align4_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX940-NEXT: v_min_f16_e32 v2, v2, v4
+; GFX940-NEXT: v_and_or_b32 v2, v3, s2, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB29_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b__align4_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX11-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f16_e32 v2, v2, v4
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB29_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b__align4_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX10-NEXT: v_min_f16_e32 v2, v2, v4
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB29_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b__align4_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX90A-NEXT: v_min_f16_e32 v2, v2, v4
+; GFX90A-NEXT: v_and_or_b32 v2, v3, s6, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB29_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b__align4_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX908-NEXT: v_min_f16_e32 v2, v2, v4
+; GFX908-NEXT: v_and_or_b32 v2, v3, s6, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB29_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b__align4_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v4, v2, v2
+; GFX8-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_max_f16_e32 v2, v3, v3
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX8-NEXT: v_min_f16_e32 v2, v2, v4
+; GFX8-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB29_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_f16__offset12b__align4_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v2
+; GFX7-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB29_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %unused = atomicrmw fmin ptr %gep, half %val syncscope("agent") seq_cst, align 4
+ ret void
+}
+
+define half @flat_system_atomic_fmin_ret_f16__offset12b_pos(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fmin_ret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_max_num_f16_e32 v5, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f16_e32 v5, v5, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB30_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmin_ret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX940-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX940-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB30_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmin_ret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB30_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmin_ret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX10-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB30_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmin_ret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_e32 v5, v3, v7
+; GFX90A-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX90A-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB30_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmin_ret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX908-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX908-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB30_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmin_ret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v5
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB30_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmin_ret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v2, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_min_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v2, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB30_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %result = atomicrmw fmin ptr %gep, half %val seq_cst
+ ret half %result
+}
+
+define void @flat_system_atomic_fmin_noret_f16__offset12b_pos(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fmin_noret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_max_num_f16_e32 v6, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v2, v2, v2
+; GFX12-NEXT: v_min_num_f16_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB31_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmin_noret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX940-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX940-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX940-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB31_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmin_noret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX11-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB31_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmin_noret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX10-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX10-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB31_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmin_noret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX90A-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX90A-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX90A-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB31_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmin_noret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX908-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX908-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX908-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB31_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmin_noret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_e32 v6, v2, v2
+; GFX8-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: v_min_f16_e32 v2, v2, v6
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB31_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmin_noret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB31_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %unused = atomicrmw fmin ptr %gep, half %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; bfloat
+; --------------------------------------------------------------------
+
+define bfloat @flat_agent_atomic_fmin_ret_bf16(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB32_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB32_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB32_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB32_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB32_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB32_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB32_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB32_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr %ptr, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+define bfloat @flat_agent_atomic_fmin_ret_bf16__offset12b_pos(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB33_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB33_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB33_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB33_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB33_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB33_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB33_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB33_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %result = atomicrmw fmin ptr %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+define bfloat @flat_agent_atomic_fmin_ret_bf16__offset12b_neg(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB34_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB34_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB34_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB34_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB34_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB34_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB34_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB34_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 -1024
+ %result = atomicrmw fmin ptr %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+ }
+
+define void @flat_agent_atomic_fmin_noret_bf16(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v4, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB35_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v4, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_min_f32_e32 v4, v4, v2
+; GFX940-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX940-NEXT: v_add3_u32 v7, v7, v4, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v4, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB35_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v4, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB35_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB35_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_min_f32_e32 v4, v4, v2
+; GFX90A-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX90A-NEXT: v_add3_u32 v7, v7, v4, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB35_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB35_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX8-NEXT: v_bfe_u32 v8, v3, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v3
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB35_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v3, 0xffff, v5
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v6, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB35_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmin ptr %ptr, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmin_noret_bf16__offset12b_pos(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB36_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB36_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB36_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB36_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB36_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB36_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB36_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB36_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %unused = atomicrmw fmin ptr %gep, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmin_noret_bf16__offset12b_neg(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB37_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB37_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB37_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB37_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB37_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB37_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB37_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB37_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 -1024
+ %unused = atomicrmw fmin ptr %gep, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define bfloat @flat_agent_atomic_fmin_ret_bf16__offset12b_pos__align4(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_pos__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB38_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_pos__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX940-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX940-NEXT: v_bfe_u32 v4, v3, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX940-NEXT: v_add3_u32 v4, v4, v3, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v3, v4, v6, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX940-NEXT: v_and_or_b32 v4, v5, s3, v3
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB38_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_pos__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB38_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_pos__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX10-NEXT: v_min_f32_e32 v0, v0, v1
+; GFX10-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX10-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX10-NEXT: v_and_or_b32 v5, 0xffff0000, v6, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB38_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_pos__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX90A-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX90A-NEXT: v_bfe_u32 v4, v3, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX90A-NEXT: v_add3_u32 v4, v4, v3, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e32 v3, v4, v6, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX90A-NEXT: v_and_or_b32 v4, v5, s7, v3
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB38_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_pos__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX908-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, s7, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB38_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_pos__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_min_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v0
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX8-NEXT: v_cndmask_b32_e32 v0, v5, v7, vcc
+; GFX8-NEXT: v_or_b32_sdwa v5, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB38_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_bf16__offset12b_pos__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB38_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %result = atomicrmw fmin ptr %gep, bfloat %val syncscope("agent") seq_cst, align 4
+ ret bfloat %result
+}
+
+define void @flat_agent_atomic_fmin_noret_bf16__offset12b__align4_pos(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b__align4_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB39_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b__align4_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v5, v5, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: v_and_or_b32 v2, v3, s3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB39_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b__align4_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB39_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b__align4_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB39_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b__align4_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v5, v5, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v3, s7, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB39_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b__align4_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v5, v5, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: v_and_or_b32 v2, v3, s7, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB39_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b__align4_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v2
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v6, v7, vcc
+; GFX8-NEXT: v_or_b32_sdwa v2, v5, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB39_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b__align4_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX7-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB39_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %unused = atomicrmw fmin ptr %gep, bfloat %val syncscope("agent") seq_cst, align 4
+ ret void
+}
+
+define bfloat @flat_system_atomic_fmin_ret_bf16__offset12b_pos(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB40_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB40_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB40_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB40_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB40_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB40_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB40_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmin_ret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB40_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %result = atomicrmw fmin ptr %gep, bfloat %val seq_cst
+ ret bfloat %result
+}
+
+define void @flat_system_atomic_fmin_noret_bf16__offset12b_pos(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB41_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB41_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB41_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB41_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB41_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB41_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB41_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmin_noret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_min_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB41_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %unused = atomicrmw fmin ptr %gep, bfloat %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x half>
+; --------------------------------------------------------------------
+
+define <2 x half> @flat_agent_atomic_fmin_ret_v2f16(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v3, v4, v4
+; GFX12-NEXT: v_pk_min_num_f16 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB42_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX940-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_min_f16 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB42_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX11-NEXT: v_pk_min_f16 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB42_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX10-NEXT: v_pk_min_f16 v3, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB42_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX90A-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX90A-NEXT: v_pk_min_f16 v4, v3, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB42_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX908-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX908-NEXT: v_pk_min_f16 v3, v3, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB42_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v4, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v3
+; GFX8-NEXT: v_max_f16_sdwa v3, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v6, v6
+; GFX8-NEXT: v_min_f16_sdwa v3, v3, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB42_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX7-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[6:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB42_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr %ptr, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define <2 x half> @flat_agent_atomic_fmin_ret_v2f16__offset12b_pos(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v3, v4, v4
+; GFX12-NEXT: v_pk_min_num_f16 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB43_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX940-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_min_f16 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB43_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX11-NEXT: v_pk_min_f16 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB43_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_pk_max_f16 v1, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_pk_max_f16 v0, v6, v6
+; GFX10-NEXT: v_pk_min_f16 v5, v0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB43_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX90A-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX90A-NEXT: v_pk_min_f16 v4, v3, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB43_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX908-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX908-NEXT: v_pk_min_f16 v3, v3, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB43_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v1, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_max_f16_sdwa v0, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v6, v6
+; GFX8-NEXT: v_min_f16_sdwa v0, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB43_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v1, v[4:5]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v0
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[4:5], v[6:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB43_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 511
+ %result = atomicrmw fmin ptr %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define <2 x half> @flat_agent_atomic_fmin_ret_v2f16__offset12b_neg(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_v2f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v3, v4, v4
+; GFX12-NEXT: v_pk_min_num_f16 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB44_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_v2f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v4
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v5, vcc
+; GFX940-NEXT: flat_load_dword v0, v[0:1]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v1, v2, v2
+; GFX940-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_max_f16 v0, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_min_f16 v2, v0, v1
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v0, v[4:5], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB44_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_v2f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: flat_load_b32 v0, v[4:5]
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_pk_max_f16 v1, v2, v2
+; GFX11-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v0, v6, v6
+; GFX11-NEXT: v_pk_min_f16 v5, v0, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v0, v[3:4], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB44_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_v2f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_pk_max_f16 v1, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_pk_max_f16 v0, v6, v6
+; GFX10-NEXT: v_pk_min_f16 v5, v0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB44_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_v2f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v0, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v1, v2, v2
+; GFX90A-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_max_f16 v0, v3, v3
+; GFX90A-NEXT: v_pk_min_f16 v2, v0, v1
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB44_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_v2f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v0, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v1, v2, v2
+; GFX908-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v0
+; GFX908-NEXT: v_pk_max_f16 v0, v6, v6
+; GFX908-NEXT: v_pk_min_f16 v5, v0, v1
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB44_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_v2f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v1, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_max_f16_sdwa v0, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v6, v6
+; GFX8-NEXT: v_min_f16_sdwa v0, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB44_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_v2f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v1, v[4:5]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v0
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[4:5], v[6:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB44_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 -512
+ %result = atomicrmw fmin ptr %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define void @flat_agent_atomic_fmin_noret_v2f16(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_pk_max_num_f16 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_min_num_f16 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB45_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX940-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB45_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB45_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX10-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB45_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX90A-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX90A-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB45_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX908-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX908-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB45_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v4, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v2, v2
+; GFX8-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v2, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v6, v3, v3
+; GFX8-NEXT: v_min_f16_sdwa v2, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v6, v6, v5
+; GFX8-NEXT: v_or_b32_e32 v2, v6, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB45_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: flat_atomic_cmpswap v7, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB45_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmin ptr %ptr, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmin_noret_v2f16__offset12b_pos(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_pk_max_num_f16 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_min_num_f16 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB46_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX940-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB46_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB46_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX10-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB46_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX90A-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX90A-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB46_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX908-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX908-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB46_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v4, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v2, v2
+; GFX8-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v2, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v6, v3, v3
+; GFX8-NEXT: v_min_f16_sdwa v2, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v6, v6, v5
+; GFX8-NEXT: v_or_b32_e32 v2, v6, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB46_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: flat_atomic_cmpswap v7, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB46_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 511
+ %unused = atomicrmw fmin ptr %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmin_noret_v2f16__offset12b_neg(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_v2f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: v_pk_max_num_f16 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_min_num_f16 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB47_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_v2f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX940-NEXT: flat_load_dword v3, v[4:5]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX940-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB47_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_v2f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: flat_load_b32 v3, v[3:4]
+; GFX11-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB47_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_v2f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX10-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB47_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_v2f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v1, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX90A-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v0, v1, v1
+; GFX90A-NEXT: v_pk_min_f16 v0, v0, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB47_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_v2f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v1, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX908-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v0, v1, v1
+; GFX908-NEXT: v_pk_min_f16 v0, v0, v2
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB47_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_v2f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v4, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v2, v2
+; GFX8-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v2, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v6, v3, v3
+; GFX8-NEXT: v_min_f16_sdwa v2, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v6, v6, v5
+; GFX8-NEXT: v_or_b32_e32 v2, v6, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB47_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_v2f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: flat_atomic_cmpswap v7, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB47_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 -512
+ %unused = atomicrmw fmin ptr %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x half> @flat_system_atomic_fmin_ret_v2f16__offset12b_pos(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_pk_max_num_f16 v2, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v3, v4, v4
+; GFX12-NEXT: v_pk_min_num_f16 v3, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB48_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX940-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_min_f16 v4, v3, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB48_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX11-NEXT: v_pk_min_f16 v3, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB48_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_pk_max_f16 v1, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_pk_max_f16 v0, v6, v6
+; GFX10-NEXT: v_pk_min_f16 v5, v0, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB48_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX90A-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_max_f16 v3, v5, v5
+; GFX90A-NEXT: v_pk_min_f16 v4, v3, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB48_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v2, v2, v2
+; GFX908-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_max_f16 v3, v4, v4
+; GFX908-NEXT: v_pk_min_f16 v3, v3, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB48_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v1, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v2, v2, v2
+; GFX8-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_max_f16_sdwa v0, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v6, v6
+; GFX8-NEXT: v_min_f16_sdwa v0, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB48_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmin_ret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v1, v[4:5]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v0
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[4:5], v[6:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB48_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 511
+ %result = atomicrmw fmin ptr %gep, <2 x half> %val seq_cst
+ ret <2 x half> %result
+}
+
+define void @flat_system_atomic_fmin_noret_v2f16__offset12b_pos(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_pk_max_num_f16 v4, v2, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v2, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_min_num_f16 v2, v2, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB49_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX940-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB49_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB49_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX10-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB49_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX90A-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX90A-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB49_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_pk_max_f16 v4, v2, v2
+; GFX908-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v2, v3, v3
+; GFX908-NEXT: v_pk_min_f16 v2, v2, v4
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB49_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_max_f16_sdwa v4, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v2, v2
+; GFX8-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v2, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v6, v3, v3
+; GFX8-NEXT: v_min_f16_sdwa v2, v2, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v6, v6, v5
+; GFX8-NEXT: v_or_b32_e32 v2, v6, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB49_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmin_noret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: flat_atomic_cmpswap v7, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB49_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 511
+ %unused = atomicrmw fmin ptr %gep, <2 x half> %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x bfloat>
+; --------------------------------------------------------------------
+
+define <2 x bfloat> @flat_agent_atomic_fmin_ret_v2bf16(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_min_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB50_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB50_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB50_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB50_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB50_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB50_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v3
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v3, 16
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB50_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v2, v2, v7, 16
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB50_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fmin ptr %ptr, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define <2 x bfloat> @flat_agent_atomic_fmin_ret_v2bf16__offset12b_pos(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_min_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB51_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB51_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB51_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_min_f32_e32 v0, v0, v1
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v0
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v0, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v0, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB51_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB51_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB51_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_min_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB51_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[4:5]
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v0, v0, v7, 16
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB51_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 511
+ %result = atomicrmw fmin ptr %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define <2 x bfloat> @flat_agent_atomic_fmin_ret_v2bf16__offset12b_neg(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_ret_v2bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_min_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB52_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_ret_v2bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v4
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v5, vcc
+; GFX940-NEXT: flat_load_dword v0, v[0:1]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v3, 0xffff0000, v7
+; GFX940-NEXT: v_min_f32_e32 v0, v0, v1
+; GFX940-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX940-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v3, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v3
+; GFX940-NEXT: v_add3_u32 v6, v6, v0, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v3, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v0, v0
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v0, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v3, v0, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v0, v[4:5], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB52_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_ret_v2bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: flat_load_b32 v0, v[4:5]
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_min_f32_e32 v0, v0, v1
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v0
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v0, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v0, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v0, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v0, v[3:4], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB52_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_ret_v2bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_min_f32_e32 v0, v0, v1
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v0
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v0, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v0, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB52_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_ret_v2bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v0, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v3, 0xffff0000, v7
+; GFX90A-NEXT: v_min_f32_e32 v0, v0, v1
+; GFX90A-NEXT: v_min_f32_e32 v3, v3, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v3, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v3
+; GFX90A-NEXT: v_add3_u32 v6, v6, v0, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v3, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX90A-NEXT: v_cndmask_b32_e64 v0, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v3, v0, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB52_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_ret_v2bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v0, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_min_f32_e32 v0, v0, v1
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v0, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX908-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v0, s9
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB52_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_ret_v2bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_min_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB52_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_ret_v2bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[4:5]
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v0, v0, v7, 16
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB52_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 -512
+ %result = atomicrmw fmin ptr %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @flat_agent_atomic_fmin_noret_v2bf16(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_min_num_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB53_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB53_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB53_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB53_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB53_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB53_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB53_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX7-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB53_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fmin ptr %ptr, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmin_noret_v2bf16__offset12b_pos(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_min_num_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB54_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB54_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB54_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB54_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB54_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB54_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB54_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX7-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB54_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 511
+ %unused = atomicrmw fmin ptr %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fmin_noret_v2bf16__offset12b_neg(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fmin_noret_v2bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_min_num_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB55_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fmin_noret_v2bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX940-NEXT: flat_load_dword v3, v[4:5]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB55_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fmin_noret_v2bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: flat_load_b32 v3, v[3:4]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB55_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fmin_noret_v2bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB55_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fmin_noret_v2bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v1, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX90A-NEXT: v_min_f32_e32 v0, v0, v3
+; GFX90A-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX90A-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v0, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX90A-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v0, v6, v0, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB55_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fmin_noret_v2bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v1, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX908-NEXT: v_min_f32_e32 v0, v0, v5
+; GFX908-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX908-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v0, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX908-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v0, v6, v0, s9
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB55_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fmin_noret_v2bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB55_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fmin_noret_v2bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX7-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB55_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 -512
+ %unused = atomicrmw fmin ptr %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x bfloat> @flat_system_atomic_fmin_ret_v2bf16__offset12b_pos(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_min_num_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_min_num_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB56_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB56_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB56_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_min_f32_e32 v0, v0, v1
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v0
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v0, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v0, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB56_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB56_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_min_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB56_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_min_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB56_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmin_ret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[4:5]
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v0, v0, v7, 16
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB56_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 511
+ %result = atomicrmw fmin ptr %gep, <2 x bfloat> %val seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @flat_system_atomic_fmin_noret_v2bf16__offset12b_pos(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_min_num_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_min_num_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB57_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB57_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB57_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB57_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB57_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB57_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_min_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_min_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB57_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fmin_noret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX7-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB57_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 511
+ %unused = atomicrmw fmin ptr %gep, <2 x bfloat> %val seq_cst
+ ret void
+}
+
+attributes #0 = { nounwind "amdgpu-unsafe-fp-atomics"="true" }
+attributes #1 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" "amdgpu-unsafe-fp-atomics"="true" }
diff --git a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fsub.ll b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fsub.ll
new file mode 100644
index 0000000000000..bce522c0c3f0a
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fsub.ll
@@ -0,0 +1,15075 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 < %s | FileCheck -check-prefix=GFX12 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx940 < %s | FileCheck -check-prefix=GFX940 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx90a < %s | FileCheck -check-prefix=GFX90A %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx908 < %s | FileCheck -check-prefix=GFX908 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tonga < %s | FileCheck -check-prefix=GFX8 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=hawaii < %s | FileCheck -check-prefix=GFX7 %s
+
+; --------------------------------------------------------------------
+; float
+; --------------------------------------------------------------------
+
+define float @flat_agent_atomic_fsub_ret_f32(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB0_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB0_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB0_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB0_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB0_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB0_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB0_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB0_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr %ptr, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @flat_agent_atomic_fsub_ret_f32__offset12b_pos(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB1_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB1_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB1_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB1_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB1_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB1_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB1_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v1, v0
+; GFX7-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB1_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %result = atomicrmw fsub ptr %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @flat_agent_atomic_fsub_ret_f32__offset12b_neg(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB2_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v4
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v5, vcc
+; GFX940-NEXT: flat_load_dword v0, v[0:1]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, v0
+; GFX940-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB2_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: flat_load_b32 v0, v[4:5]
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v1, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v0, v[3:4], v[0:1] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB2_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB2_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v0, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB2_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v0, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB2_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB2_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB2_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v1, v0
+; GFX7-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 -512
+ %result = atomicrmw fsub ptr %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define void @flat_agent_atomic_fsub_noret_f32(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_f32:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v4, v[0:1]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB3_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_f32:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB3_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v4, v[0:1]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB3_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_f32:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB3_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_f32:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB3_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_f32:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB3_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_f32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB3_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_f32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB3_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fsub ptr %ptr, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fsub_noret_f32__offset12b_pos(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v4, v[0:1] offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB4_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v5, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB4_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v4, v[0:1] offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB4_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB4_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v5, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB4_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v4, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB4_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB4_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB4_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %unused = atomicrmw fsub ptr %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fsub_noret_f32__offset12b_neg(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v4, v[0:1] offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB5_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX940-NEXT: flat_load_dword v5, v[4:5]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB5_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: flat_load_b32 v4, v[3:4]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB5_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB5_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v1, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB5_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v1, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB5_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 -512
+ %unused = atomicrmw fsub ptr %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define float @flat_system_atomic_fsub_ret_f32__offset12b_pos(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fsub_ret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB6_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fsub_ret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB6_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fsub_ret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB6_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fsub_ret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB6_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fsub_ret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB6_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fsub_ret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB6_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fsub_ret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB6_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fsub_ret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v1, v0
+; GFX7-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB6_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %result = atomicrmw fsub ptr %gep, float %val seq_cst
+ ret float %result
+}
+
+define void @flat_system_atomic_fsub_noret_f32__offset12b_pos(ptr %ptr, float %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fsub_noret_f32__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v4, v[0:1] offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB7_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fsub_noret_f32__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v5, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB7_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fsub_noret_f32__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v4, v[0:1] offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB7_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fsub_noret_f32__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB7_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fsub_noret_f32__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v5, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB7_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fsub_noret_f32__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v4, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB7_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fsub_noret_f32__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB7_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fsub_noret_f32__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB7_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %unused = atomicrmw fsub ptr %gep, float %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; float with ftz/daz
+; --------------------------------------------------------------------
+
+define float @flat_agent_atomic_fsub_ret_f32__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_f32__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB8_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_f32__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB8_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_f32__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB8_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_f32__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB8_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_f32__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB8_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_f32__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_f32__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_f32__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB8_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB8_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr %ptr, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @flat_agent_atomic_fsub_ret_f32__offset12b_pos__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB9_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB9_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB9_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB9_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB9_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB9_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB9_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v1, v0
+; GFX7-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB9_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %result = atomicrmw fsub ptr %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define float @flat_agent_atomic_fsub_ret_f32__offset12b_neg__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_neg__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB10_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_neg__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v4
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v5, vcc
+; GFX940-NEXT: flat_load_dword v0, v[0:1]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, v0
+; GFX940-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB10_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_neg__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: flat_load_b32 v0, v[4:5]
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v1, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v0, v[3:4], v[0:1] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB10_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_neg__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB10_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_neg__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v0, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB10_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_neg__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v0, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB10_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_neg__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB10_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_f32__offset12b_neg__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v1, v0
+; GFX7-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB10_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 -512
+ %result = atomicrmw fsub ptr %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define void @flat_agent_atomic_fsub_noret_f32__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_f32__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v4, v[0:1]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB11_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_f32__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB11_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_f32__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v4, v[0:1]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB11_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_f32__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB11_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_f32__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_f32__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_f32__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_f32__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB11_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB11_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fsub ptr %ptr, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fsub_noret_f32__offset12b_pos__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v4, v[0:1] offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB12_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v5, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB12_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v4, v[0:1] offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB12_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB12_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v5, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB12_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v4, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB12_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB12_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB12_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %unused = atomicrmw fsub ptr %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fsub_noret_f32__offset12b_neg__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_neg__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v4, v[0:1] offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB13_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_neg__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX940-NEXT: flat_load_dword v5, v[4:5]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB13_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_neg__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: flat_load_b32 v4, v[3:4]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB13_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_neg__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB13_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_neg__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v1, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB13_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_neg__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v1, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB13_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_neg__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB13_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_f32__offset12b_neg__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB13_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 -512
+ %unused = atomicrmw fsub ptr %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define float @flat_system_atomic_fsub_ret_f32__offset12b_pos__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_system_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB14_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB14_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB14_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB14_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fsub_ret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[3:4]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB14_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v1, v0
+; GFX7-NEXT: v_sub_f32_e32 v0, v1, v2
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %result = atomicrmw fsub ptr %gep, float %val seq_cst
+ ret float %result
+}
+
+define void @flat_system_atomic_fsub_noret_f32__offset12b_pos__ftz(ptr %ptr, float %val) #1 {
+; GFX12-LABEL: flat_system_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v4, v[0:1] offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB15_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v5, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB15_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v4, v[0:1] offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB15_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB15_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v5, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_sub_f32_e32 v4, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB15_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v4, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB15_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB15_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fsub_noret_f32__offset12b_pos__ftz:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_sub_f32_e32 v3, v4, v2
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB15_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr %ptr, i64 511
+ %unused = atomicrmw fsub ptr %gep, float %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; double
+; --------------------------------------------------------------------
+
+define double @flat_agent_atomic_fsub_ret_f64(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b64 v[4:5], v[0:1]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_add_f64_e64 v[4:5], v[6:7], -v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB16_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB16_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b64 v[4:5], v[0:1]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB16_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB16_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX90A-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB16_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dwordx2 v[4:5], v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB16_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 4, v0
+; GFX8-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: flat_load_dword v5, v[5:6]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v7, v5
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB16_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: v_mov_b32_e32 v1, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v5, vcc, 4, v0
+; GFX7-NEXT: v_addc_u32_e32 v6, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: flat_load_dword v5, v[5:6]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB16_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v4
+; GFX7-NEXT: v_mov_b32_e32 v1, v5
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr %ptr, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define double @flat_agent_atomic_fsub_ret_f64__offset12b_pos(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_f64__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b64 v[4:5], v[0:1] offset:2040
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_add_f64_e64 v[4:5], v[6:7], -v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] offset:2040 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB17_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_f64__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dwordx2 v[4:5], v[0:1] offset:2040
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] offset:2040 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB17_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_f64__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b64 v[4:5], v[0:1] offset:2040
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] offset:2040 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB17_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_f64__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7f8, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dwordx2 v[0:1], v[4:5]
+; GFX10-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v9, v1
+; GFX10-NEXT: v_mov_b32_e32 v8, v0
+; GFX10-NEXT: v_add_f64 v[6:7], v[8:9], -v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[8:9]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB17_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_f64__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dwordx2 v[4:5], v[0:1] offset:2040
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX90A-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] offset:2040 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB17_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_f64__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dwordx2 v[4:5], v[0:1] offset:2040
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] offset:2040 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_f64__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7f8, v0
+; GFX8-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v1, v[0:1]
+; GFX8-NEXT: flat_load_dword v0, v[4:5]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v9, v1
+; GFX8-NEXT: v_mov_b32_e32 v8, v0
+; GFX8-NEXT: v_add_f64 v[6:7], v[8:9], -v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_f64__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7f8, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v1, v[0:1]
+; GFX7-NEXT: flat_load_dword v0, v[4:5]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB17_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v9, v1
+; GFX7-NEXT: v_mov_b32_e32 v8, v0
+; GFX7-NEXT: v_add_f64 v[6:7], v[8:9], -v[2:3]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr %ptr, i64 255
+ %result = atomicrmw fsub ptr %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define double @flat_agent_atomic_fsub_ret_f64__offset12b_neg(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_f64__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b64 v[4:5], v[0:1] offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_add_f64_e64 v[4:5], v[6:7], -v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB18_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_f64__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v4
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v5, vcc
+; GFX940-NEXT: flat_load_dwordx2 v[0:1], v[0:1]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[8:9], v[0:1]
+; GFX940-NEXT: v_add_f64 v[6:7], v[8:9], -v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB18_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_f64__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v5, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v5
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v4, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v5, vcc_lo, 0xfffff800, v5
+; GFX11-NEXT: v_add_co_ci_u32_e32 v6, vcc_lo, -1, v4, vcc_lo
+; GFX11-NEXT: flat_load_b64 v[0:1], v[0:1]
+; GFX11-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v10, v1 :: v_dual_mov_b32 v9, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_f64 v[7:8], v[9:10], -v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[0:1], v[5:6], v[7:10] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[9:10]
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB18_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_f64__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dwordx2 v[0:1], v[4:5]
+; GFX10-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v9, v1
+; GFX10-NEXT: v_mov_b32_e32 v8, v0
+; GFX10-NEXT: v_add_f64 v[6:7], v[8:9], -v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[8:9]
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB18_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_f64__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dwordx2 v[0:1], v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[8:9], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: v_add_f64 v[6:7], v[8:9], -v[2:3]
+; GFX90A-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB18_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_f64__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dwordx2 v[0:1], v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v9, v1
+; GFX908-NEXT: v_mov_b32_e32 v8, v0
+; GFX908-NEXT: v_add_f64 v[6:7], v[8:9], -v[2:3]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB18_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_f64__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff804, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v1, v[0:1]
+; GFX8-NEXT: flat_load_dword v0, v[4:5]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v9, v1
+; GFX8-NEXT: v_mov_b32_e32 v8, v0
+; GFX8-NEXT: v_add_f64 v[6:7], v[8:9], -v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB18_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_f64__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff804, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v1, v[0:1]
+; GFX7-NEXT: flat_load_dword v0, v[4:5]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v9, v1
+; GFX7-NEXT: v_mov_b32_e32 v8, v0
+; GFX7-NEXT: v_add_f64 v[6:7], v[8:9], -v[2:3]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[6:9] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[8:9]
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB18_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr %ptr, i64 -256
+ %result = atomicrmw fsub ptr %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define void @flat_agent_atomic_fsub_noret_f64(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_f64:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b64 v[6:7], v[0:1]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_add_f64_e64 v[4:5], v[6:7], -v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB19_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_f64:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dwordx2 v[6:7], v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB19_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_f64:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b64 v[6:7], v[0:1]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB19_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_f64:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dwordx2 v[6:7], v[0:1]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB19_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_f64:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dwordx2 v[6:7], v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX90A-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB19_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_f64:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dwordx2 v[6:7], v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB19_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_f64:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 4, v0
+; GFX8-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v6, v[0:1]
+; GFX8-NEXT: flat_load_dword v7, v[4:5]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v7, v5
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB19_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_f64:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 4, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v6, v[0:1]
+; GFX7-NEXT: flat_load_dword v7, v[4:5]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB19_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB19_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fsub ptr %ptr, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fsub_noret_f64__offset12b_pos(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_f64__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b64 v[6:7], v[0:1] offset:2040
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_add_f64_e64 v[4:5], v[6:7], -v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] offset:2040 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB20_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_f64__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dwordx2 v[6:7], v[0:1] offset:2040
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] offset:2040 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB20_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_f64__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b64 v[6:7], v[0:1] offset:2040
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] offset:2040 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB20_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_f64__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7f8, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dwordx2 v[6:7], v[0:1]
+; GFX10-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB20_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_f64__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dwordx2 v[6:7], v[0:1] offset:2040
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX90A-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] offset:2040 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB20_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_f64__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dwordx2 v[6:7], v[0:1] offset:2040
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] offset:2040 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v7, v5
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB20_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_f64__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7f8, v0
+; GFX8-NEXT: v_addc_u32_e32 v9, vcc, 0, v1, vcc
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v7, v[0:1]
+; GFX8-NEXT: flat_load_dword v6, v[8:9]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[8:9], v[4:7] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v7, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB20_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_f64__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v8, vcc, 0x7f8, v0
+; GFX7-NEXT: v_addc_u32_e32 v9, vcc, 0, v1, vcc
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v7, v[0:1]
+; GFX7-NEXT: flat_load_dword v6, v[8:9]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB20_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[8:9], v[4:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v7, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB20_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr %ptr, i64 255
+ %unused = atomicrmw fsub ptr %gep, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fsub_noret_f64__offset12b_neg(ptr %ptr, double %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_f64__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b64 v[6:7], v[0:1] offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_add_f64_e64 v[4:5], v[6:7], -v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX12-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB21_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_f64__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX940-NEXT: flat_load_dwordx2 v[6:7], v[4:5]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[6:7]
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB21_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_f64__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: flat_load_b64 v[6:7], v[4:5]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b64 v[4:5], v[0:1], v[4:7] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX11-NEXT: v_dual_mov_b32 v7, v5 :: v_dual_mov_b32 v6, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB21_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_f64__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dwordx2 v[6:7], v[0:1]
+; GFX10-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap_x2 v[4:5], v[0:1], v[4:7] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[4:5], v[6:7]
+; GFX10-NEXT: v_mov_b32_e32 v7, v5
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB21_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_f64__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v8, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v9, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dwordx2 v[6:7], v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX90A-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[8:9], v[4:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[6:7]
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB21_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_f64__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v8, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v9, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dwordx2 v[6:7], v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX908-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[8:9], v[4:7] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v7, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v6, v0
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB21_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_f64__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v9, vcc, -1, v1, vcc
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff804, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v7, v[0:1]
+; GFX8-NEXT: flat_load_dword v6, v[8:9]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX8-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[8:9], v[4:7] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v7, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB21_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_f64__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v8, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v9, vcc, -1, v1, vcc
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff804, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v7, v[0:1]
+; GFX7-NEXT: flat_load_dword v6, v[8:9]
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB21_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_f64 v[4:5], v[6:7], -v[2:3]
+; GFX7-NEXT: flat_atomic_cmpswap_x2 v[0:1], v[8:9], v[4:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v7, v1
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v6, v0
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB21_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr %ptr, i64 -256
+ %unused = atomicrmw fsub ptr %gep, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; half
+; --------------------------------------------------------------------
+
+define half @flat_agent_atomic_fsub_ret_f16(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB22_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: flat_load_dword v4, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX940-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB22_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB22_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB22_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v4, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX90A-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB22_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v4
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX908-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX908-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX908-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB22_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB22_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: .LBB22_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v2, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_sub_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v2, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB22_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr %ptr, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+define half @flat_agent_atomic_fsub_ret_f16__offset12b_pos(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB23_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[6:7], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v6
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: flat_load_dword v4, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX940-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB23_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB23_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB23_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v4, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX90A-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB23_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v4
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX908-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX908-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX908-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB23_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB23_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: .LBB23_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v2, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_sub_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v2, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB23_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %result = atomicrmw fsub ptr %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+define half @flat_agent_atomic_fsub_ret_f16__offset12b_neg(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB24_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[6:7], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v6
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: flat_load_dword v4, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX940-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB24_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB24_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB24_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v4, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX90A-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB24_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v4
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX908-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX908-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX908-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB24_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB24_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: .LBB24_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v2, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_sub_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v2, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB24_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 -1024
+ %result = atomicrmw fsub ptr %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+ }
+
+define void @flat_agent_atomic_fsub_noret_f16(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v3, v0
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v4, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB25_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX940-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB25_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v4, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB25_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX10-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB25_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX90A-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB25_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX908-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB25_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX8-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB25_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: v_lshl_b32_e32 v3, 0xffff, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_not_b32_e32 v6, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: .LBB25_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v2, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX7-NEXT: v_sub_f32_e32 v3, v3, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v2, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB25_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fsub ptr %ptr, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fsub_noret_f16__offset12b_pos(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v4, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB26_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX940-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB26_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v4, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB26_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX10-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB26_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX90A-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB26_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX908-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB26_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX8-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB26_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB26_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB26_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %unused = atomicrmw fsub ptr %gep, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fsub_noret_f16__offset12b_neg(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v4, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB27_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX940-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB27_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v4, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB27_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX10-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB27_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX90A-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB27_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX908-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB27_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX8-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB27_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB27_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB27_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 -1024
+ %unused = atomicrmw fsub ptr %gep, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define half @flat_agent_atomic_fsub_ret_f16__offset12b_pos__align4(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_pos__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f16_e32 v3, v4, v2
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB28_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_pos__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_sub_f16_e32 v3, v5, v2
+; GFX940-NEXT: v_and_or_b32 v4, v5, s2, v3
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB28_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_pos__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f16_e32 v3, v4, v2
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB28_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_pos__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_sub_f16_e32 v0, v1, v2
+; GFX10-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX10-NEXT: v_and_or_b32 v0, 0xffff0000, v1, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB28_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_pos__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_sub_f16_e32 v3, v5, v2
+; GFX90A-NEXT: v_and_or_b32 v4, v5, s6, v3
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB28_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_pos__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_sub_f16_e32 v3, v4, v2
+; GFX908-NEXT: v_and_or_b32 v3, v4, s6, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB28_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_pos__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_sub_f16_e32 v0, v1, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX8-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB28_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_f16__offset12b_pos__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: .LBB28_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB28_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %result = atomicrmw fsub ptr %gep, half %val syncscope("agent") seq_cst, align 4
+ ret half %result
+}
+
+define void @flat_agent_atomic_fsub_noret_f16__offset12b__align4_pos(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b__align4_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v4, v[0:1] offset:2046
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_sub_f16_e32 v3, v4, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB29_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b__align4_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v5, v[0:1] offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: s_mov_b32 s2, 0xffff0000
+; GFX940-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_sub_f16_e32 v3, v5, v2
+; GFX940-NEXT: v_and_or_b32 v4, v5, s2, v3
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB29_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b__align4_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v4, v[0:1] offset:2046
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_sub_f16_e32 v3, v4, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB29_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b__align4_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_sub_f16_e32 v3, v4, v2
+; GFX10-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX10-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB29_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b__align4_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v5, v[0:1] offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX90A-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_sub_f16_e32 v3, v5, v2
+; GFX90A-NEXT: v_and_or_b32 v4, v5, s6, v3
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB29_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b__align4_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v4, v[0:1] offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: s_mov_b32 s6, 0xffff0000
+; GFX908-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_sub_f16_e32 v3, v4, v2
+; GFX908-NEXT: v_and_or_b32 v3, v4, s6, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB29_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b__align4_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_sub_f16_e32 v3, v4, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB29_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_f16__offset12b__align4_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v2
+; GFX7-NEXT: .LBB29_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB29_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %unused = atomicrmw fsub ptr %gep, half %val syncscope("agent") seq_cst, align 4
+ ret void
+}
+
+define half @flat_system_atomic_fsub_ret_f16__offset12b_pos(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fsub_ret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB30_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fsub_ret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[6:7], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v6
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: flat_load_dword v4, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v6
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX940-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB30_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fsub_ret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB30_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fsub_ret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX10-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB30_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fsub_ret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v4, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX90A-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB30_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fsub_ret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v7, v4
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, v3, v7
+; GFX908-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX908-NEXT: v_and_or_b32 v6, v7, v5, v4
+; GFX908-NEXT: flat_atomic_cmpswap v4, v[0:1], v[6:7] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB30_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fsub_ret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX8-NEXT: v_sub_f16_e32 v5, v5, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB30_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fsub_ret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v2
+; GFX7-NEXT: v_and_b32_e32 v2, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: .LBB30_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v2, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_sub_f32_e32 v5, v5, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v2, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB30_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %result = atomicrmw fsub ptr %gep, half %val seq_cst
+ ret half %result
+}
+
+define void @flat_system_atomic_fsub_noret_f16__offset12b_pos(ptr %ptr, half %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fsub_noret_f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v4, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB31_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fsub_noret_f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX940-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB31_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fsub_noret_f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v4, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB31_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fsub_noret_f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX10-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB31_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fsub_noret_f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v3, v5
+; GFX90A-NEXT: v_sub_f16_e32 v4, v4, v2
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v3, v4
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB31_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fsub_noret_f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX908-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB31_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fsub_noret_f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX8-NEXT: v_sub_f16_e32 v3, v3, v2
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB31_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fsub_noret_f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v2
+; GFX7-NEXT: v_lshl_b32_e32 v2, 0xffff, v4
+; GFX7-NEXT: v_not_b32_e32 v6, v2
+; GFX7-NEXT: .LBB31_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v6
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB31_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr %ptr, i64 1023
+ %unused = atomicrmw fsub ptr %gep, half %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; bfloat
+; --------------------------------------------------------------------
+
+define bfloat @flat_agent_atomic_fsub_ret_bf16(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB32_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB32_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB32_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB32_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB32_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB32_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB32_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB32_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB32_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr %ptr, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+define bfloat @flat_agent_atomic_fsub_ret_bf16__offset12b_pos(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB33_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB33_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB33_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB33_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB33_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB33_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB33_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB33_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB33_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %result = atomicrmw fsub ptr %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+define bfloat @flat_agent_atomic_fsub_ret_bf16__offset12b_neg(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB34_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB34_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB34_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB34_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB34_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB34_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB34_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB34_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB34_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 -1024
+ %result = atomicrmw fsub ptr %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+ }
+
+define void @flat_agent_atomic_fsub_noret_bf16(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: flat_load_b32 v4, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v6, v3
+; GFX12-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB35_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v6, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v4, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_sub_f32_e32 v4, v4, v2
+; GFX940-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX940-NEXT: v_add3_u32 v7, v7, v4, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v4, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB35_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: flat_load_b32 v4, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v6, v3
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB35_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v3, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v6, v3
+; GFX10-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB35_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v6, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_sub_f32_e32 v4, v4, v2
+; GFX90A-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX90A-NEXT: v_add3_u32 v7, v7, v4, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v4, v5, v6, v4
+; GFX90A-NEXT: flat_atomic_cmpswap v4, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB35_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v3, v0
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX908-NEXT: v_not_b32_e32 v6, v3
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v3, v4, v6, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB35_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v3, v0
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v3, v5, s4
+; GFX8-NEXT: v_not_b32_e32 v6, v3
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v3, v5, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX8-NEXT: v_bfe_u32 v8, v3, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v3
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX8-NEXT: v_lshlrev_b32_sdwa v3, v5, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB35_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v3, v0
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v4, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v3, 0xffff, v5
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v6, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB35_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, v5, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v7, v4, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v7, v3
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB35_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fsub ptr %ptr, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fsub_noret_bf16__offset12b_pos(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB36_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB36_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB36_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB36_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB36_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB36_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB36_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB36_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB36_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %unused = atomicrmw fsub ptr %gep, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fsub_noret_bf16__offset12b_neg(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB37_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB37_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB37_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB37_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB37_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB37_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB37_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB37_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB37_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 -1024
+ %unused = atomicrmw fsub ptr %gep, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define bfloat @flat_agent_atomic_fsub_ret_bf16__offset12b_pos__align4(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_pos__align4:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX12-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX12-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB38_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_pos__align4:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX940-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX940-NEXT: v_bfe_u32 v4, v3, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX940-NEXT: v_add3_u32 v4, v4, v3, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v3, v4, v6, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX940-NEXT: v_and_or_b32 v4, v5, s3, v3
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB38_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_pos__align4:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX11-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB38_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_pos__align4:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX10-NEXT: v_sub_f32_e32 v0, v0, v1
+; GFX10-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX10-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX10-NEXT: v_and_or_b32 v5, 0xffff0000, v6, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB38_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_pos__align4:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX90A-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX90A-NEXT: v_bfe_u32 v4, v3, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX90A-NEXT: v_add3_u32 v4, v4, v3, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e32 v3, v4, v6, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX90A-NEXT: v_and_or_b32 v4, v5, s7, v3
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB38_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_pos__align4:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX908-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX908-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v3
+; GFX908-NEXT: v_add3_u32 v5, v5, v3, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX908-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX908-NEXT: v_and_or_b32 v3, v4, s7, v3
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB38_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_pos__align4:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_sub_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v0
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX8-NEXT: v_cndmask_b32_e32 v0, v5, v7, vcc
+; GFX8-NEXT: v_or_b32_sdwa v5, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB38_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_bf16__offset12b_pos__align4:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB38_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v4
+; GFX7-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX7-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB38_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %result = atomicrmw fsub ptr %gep, bfloat %val syncscope("agent") seq_cst, align 4
+ ret bfloat %result
+}
+
+define void @flat_agent_atomic_fsub_noret_bf16__offset12b__align4_pos(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b__align4_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB39_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b__align4_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: s_mov_b32 s3, 0xffff0000
+; GFX940-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v5, v5, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: v_and_or_b32 v2, v3, s3, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2046 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB39_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b__align4_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2046
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB39_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b__align4_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB39_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b__align4_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX90A-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v5, v5, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: v_and_or_b32 v2, v3, s7, v2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2046 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB39_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b__align4_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2046
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: s_mov_b32 s7, 0xffff0000
+; GFX908-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_bfe_u32 v5, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v5, v5, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: v_and_or_b32 v2, v3, s7, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2046 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB39_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b__align4_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_bfe_u32 v6, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v2
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v6, v7, vcc
+; GFX8-NEXT: v_or_b32_sdwa v2, v5, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB39_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b__align4_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX7-NEXT: .LBB39_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v5, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB39_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %unused = atomicrmw fsub ptr %gep, bfloat %val syncscope("agent") seq_cst, align 4
+ ret void
+}
+
+define bfloat @flat_system_atomic_fsub_ret_bf16__offset12b_pos(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX12-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v5, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX12-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v4, v4
+; GFX12-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB40_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v3, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v4, v3, s0
+; GFX940-NEXT: v_not_b32_e32 v4, v4
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v5
+; GFX940-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB40_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX11-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v5, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX11-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v4, v4
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB40_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX10-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v5, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX10-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v4, v4
+; GFX10-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v5
+; GFX10-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v6
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB40_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX90A-NEXT: v_not_b32_e32 v4, v4
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v5
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v5, v3, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v4, v5
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v5, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v5, v7
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB40_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX908-NEXT: flat_load_dword v5, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX908-NEXT: v_not_b32_e32 v4, v4
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v5
+; GFX908-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v5, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v5, v6, v4, v5
+; GFX908-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB40_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX8-NEXT: flat_load_dword v5, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v4, v3, s4
+; GFX8-NEXT: v_not_b32_e32 v4, v4
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX8-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v5
+; GFX8-NEXT: v_lshrrev_b32_sdwa v5, v3, v6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_lshlrev_b32_sdwa v5, v3, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX8-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB40_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fsub_ret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v3, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v3
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 3, v3
+; GFX7-NEXT: v_lshl_b32_e32 v4, 0xffff, v3
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v4, v4
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: .LBB40_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v6, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v7, v6, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v5
+; GFX7-NEXT: flat_atomic_cmpswap v5, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB40_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %result = atomicrmw fsub ptr %gep, bfloat %val seq_cst
+ ret bfloat %result
+}
+
+define void @flat_system_atomic_fsub_noret_bf16__offset12b_pos(ptr %ptr, bfloat %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX12-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX12-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX12-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_not_b32_e32 v5, v5
+; GFX12-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX12-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB41_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[0:1], 0x7fe
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1]
+; GFX940-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX940-NEXT: s_mov_b32 s0, 0xffff
+; GFX940-NEXT: v_lshlrev_b32_e64 v5, v4, s0
+; GFX940-NEXT: v_not_b32_e32 v5, v5
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX940-NEXT: s_movk_i32 s2, 0x7fff
+; GFX940-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s2
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX940-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX940-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB41_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX11-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX11-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_not_b32_e32 v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX11-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB41_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX10-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX10-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff
+; GFX10-NEXT: v_not_b32_e32 v5, v5
+; GFX10-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX10-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX10-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB41_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX90A-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX90A-NEXT: s_mov_b32 s4, 0xffff
+; GFX90A-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX90A-NEXT: v_not_b32_e32 v5, v5
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX90A-NEXT: s_movk_i32 s6, 0x7fff
+; GFX90A-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX90A-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX90A-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB41_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v4, vcc, 0x7fe, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
+; GFX908-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX908-NEXT: s_mov_b32 s4, 0xffff
+; GFX908-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX908-NEXT: v_not_b32_e32 v5, v5
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX908-NEXT: s_movk_i32 s6, 0x7fff
+; GFX908-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s6
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX908-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc
+; GFX908-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX908-NEXT: v_and_or_b32 v2, v3, v5, v2
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB41_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fe, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX8-NEXT: s_mov_b32 s4, 0xffff
+; GFX8-NEXT: v_lshlrev_b32_e64 v5, v4, s4
+; GFX8-NEXT: v_not_b32_e32 v5, v5
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v2
+; GFX8-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_sdwa v2, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX8-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v2
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc
+; GFX8-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX8-NEXT: v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX8-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB41_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fsub_noret_bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fe, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: v_and_b32_e32 v0, -4, v4
+; GFX7-NEXT: flat_load_dword v3, v[0:1]
+; GFX7-NEXT: v_and_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 3, v4
+; GFX7-NEXT: v_lshl_b32_e32 v5, 0xffff, v4
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_not_b32_e32 v5, v5
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
+; GFX7-NEXT: .LBB41_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, v4, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_sub_f32_e32 v2, v2, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v7, v3, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, v4, v2
+; GFX7-NEXT: v_or_b32_e32 v2, v7, v2
+; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v3, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB41_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr %ptr, i64 1023
+ %unused = atomicrmw fsub ptr %gep, bfloat %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x half>
+; --------------------------------------------------------------------
+
+define <2 x half> @flat_agent_atomic_fsub_ret_v2f16(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB42_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB42_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB42_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB42_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB42_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB42_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: v_sub_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_sub_f16_e32 v5, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB42_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v6
+; GFX7-NEXT: .LBB42_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v2, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v2
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[6:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB42_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v1, v3
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr %ptr, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define <2 x half> @flat_agent_atomic_fsub_ret_v2f16__offset12b_pos(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB43_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB43_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB43_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_pk_add_f16 v0, v1, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB43_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB43_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB43_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_sub_f16_sdwa v0, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_sub_f16_e32 v5, v1, v2
+; GFX8-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB43_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v1, v[4:5]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: .LBB43_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v0
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[4:5], v[6:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB43_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 511
+ %result = atomicrmw fsub ptr %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define <2 x half> @flat_agent_atomic_fsub_ret_v2f16__offset12b_neg(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_v2f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB44_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_v2f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v4
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v5, vcc
+; GFX940-NEXT: flat_load_dword v0, v[0:1]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, v0
+; GFX940-NEXT: v_pk_add_f16 v0, v1, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB44_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_v2f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: flat_load_b32 v0, v[4:5]
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v1, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v0, v1, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v0, v[3:4], v[0:1] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB44_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_v2f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_pk_add_f16 v0, v1, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB44_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_v2f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v0, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: v_pk_add_f16 v0, v1, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB44_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_v2f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v0, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: v_pk_add_f16 v0, v1, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB44_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_v2f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_sub_f16_sdwa v0, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_sub_f16_e32 v5, v1, v2
+; GFX8-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB44_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_v2f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v1, v[4:5]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: .LBB44_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v0
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[4:5], v[6:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB44_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 -512
+ %result = atomicrmw fsub ptr %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define void @flat_agent_atomic_fsub_noret_v2f16(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_v2f16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v4, v[0:1]
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB45_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_v2f16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v5, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB45_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_v2f16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v4, v[0:1]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB45_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_v2f16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB45_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_v2f16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v5, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB45_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_v2f16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v4, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB45_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_v2f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_sub_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_sub_f16_e32 v5, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB45_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_v2f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB45_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: flat_atomic_cmpswap v7, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB45_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fsub ptr %ptr, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fsub_noret_v2f16__offset12b_pos(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v4, v[0:1] offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB46_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v5, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB46_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v4, v[0:1] offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB46_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB46_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v5, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB46_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v4, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB46_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_sub_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_sub_f16_e32 v5, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB46_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB46_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: flat_atomic_cmpswap v7, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB46_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 511
+ %unused = atomicrmw fsub ptr %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fsub_noret_v2f16__offset12b_neg(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_v2f16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v4, v[0:1] offset:-2048
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB47_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_v2f16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX940-NEXT: flat_load_dword v5, v[4:5]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB47_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_v2f16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: flat_load_b32 v4, v[3:4]
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB47_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_v2f16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB47_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_v2f16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v1, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v0, v1, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB47_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_v2f16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v1, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v0, v1, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB47_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_v2f16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_sub_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_sub_f16_e32 v5, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB47_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_v2f16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB47_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: flat_atomic_cmpswap v7, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB47_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 -512
+ %unused = atomicrmw fsub ptr %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x half> @flat_system_atomic_fsub_ret_v2f16__offset12b_pos(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB48_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB48_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB48_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_pk_add_f16 v0, v1, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB48_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB48_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB48_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_sub_f16_sdwa v0, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_sub_f16_e32 v5, v1, v2
+; GFX8-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB48_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fsub_ret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v1, v[4:5]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: .LBB48_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v7
+; GFX7-NEXT: v_or_b32_e32 v7, v0, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v8, v0
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[4:5], v[6:7] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v7
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB48_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 511
+ %result = atomicrmw fsub ptr %gep, <2 x half> %val seq_cst
+ ret <2 x half> %result
+}
+
+define void @flat_system_atomic_fsub_noret_v2f16__offset12b_pos(ptr %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v4, v[0:1] offset:2044
+; GFX12-NEXT: s_mov_b32 s0, 0
+; GFX12-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v3
+; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB49_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v5, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[0:1], 0
+; GFX940-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX940-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v5, v3
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB49_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v4, v[0:1] offset:2044
+; GFX11-NEXT: s_mov_b32 s0, 0
+; GFX11-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v3
+; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB49_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: flat_load_dword v4, v[0:1]
+; GFX10-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v3
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB49_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v5, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[4:5], 0
+; GFX90A-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[4:5] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
+; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: v_mov_b32_e32 v5, v3
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB49_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v4, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[4:5], 0
+; GFX908-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v3, v4, v2 neg_lo:[0,1] neg_hi:[0,1]
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB49_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v4, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[4:5], 0
+; GFX8-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_sub_f16_sdwa v3, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_sub_f16_e32 v5, v4, v2
+; GFX8-NEXT: v_or_b32_e32 v3, v5, v3
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[3:4] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
+; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB49_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fsub_noret_v2f16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v2
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: .LBB49_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: flat_atomic_cmpswap v7, v[0:1], v[5:6] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB49_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr %ptr, i64 511
+ %unused = atomicrmw fsub ptr %gep, <2 x half> %val seq_cst
+ ret void
+}
+
+; --------------------------------------------------------------------
+; <2 x bfloat>
+; --------------------------------------------------------------------
+
+define <2 x bfloat> @flat_agent_atomic_fsub_ret_v2bf16(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB50_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB50_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB50_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v3
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX10-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v3, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB50_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB50_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB50_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v3
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX8-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v3
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v3, 16
+; GFX8-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB50_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX7-NEXT: .LBB50_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v5
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v4
+; GFX7-NEXT: v_alignbit_b32 v3, v2, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v2, v2, v6, 16
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[2:3] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB50_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v3
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %result = atomicrmw fsub ptr %ptr, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define <2 x bfloat> @flat_agent_atomic_fsub_ret_v2bf16__offset12b_pos(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB51_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB51_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB51_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_sub_f32_e32 v0, v0, v1
+; GFX10-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v0
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v0, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v0, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB51_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB51_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB51_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_sub_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB51_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[4:5]
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: .LBB51_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v0, v0, v6, 16
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB51_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 511
+ %result = atomicrmw fsub ptr %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define <2 x bfloat> @flat_agent_atomic_fsub_ret_v2bf16__offset12b_neg(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_ret_v2bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB52_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_ret_v2bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v4
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v5, vcc
+; GFX940-NEXT: flat_load_dword v0, v[0:1]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v3, 0xffff0000, v7
+; GFX940-NEXT: v_sub_f32_e32 v0, v0, v1
+; GFX940-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX940-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v3, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v3
+; GFX940-NEXT: v_add3_u32 v6, v6, v0, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v3, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v0, v0
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v0, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v3, v0, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v0, v[4:5], v[6:7] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB52_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_ret_v2bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v3, v0
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: v_add_co_ci_u32_e32 v5, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v3
+; GFX11-NEXT: flat_load_b32 v0, v[4:5]
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_sub_f32_e32 v0, v0, v1
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v0
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v0, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v0, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v0, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v0, v[3:4], v[5:6] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB52_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_ret_v2bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_sub_f32_e32 v0, v0, v1
+; GFX10-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v0
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v0, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v0, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB52_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_ret_v2bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v0, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v3, 0xffff0000, v7
+; GFX90A-NEXT: v_sub_f32_e32 v0, v0, v1
+; GFX90A-NEXT: v_sub_f32_e32 v3, v3, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v3, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v3
+; GFX90A-NEXT: v_add3_u32 v6, v6, v0, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v3, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX90A-NEXT: v_cndmask_b32_e64 v0, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v3, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v3, v0, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[6:7] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB52_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_ret_v2bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v0, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_sub_f32_e32 v0, v0, v1
+; GFX908-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v0, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX908-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v0, s9
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB52_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_ret_v2bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_sub_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB52_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_ret_v2bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[4:5]
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: .LBB52_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v0, v0, v6, 16
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB52_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 -512
+ %result = atomicrmw fsub ptr %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @flat_agent_atomic_fsub_noret_v2bf16(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_v2bf16:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1]
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB53_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_v2bf16:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB53_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_v2bf16:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB53_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_v2bf16:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB53_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_v2bf16:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB53_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_v2bf16:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB53_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_v2bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB53_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_v2bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: .LBB53_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[4:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB53_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %unused = atomicrmw fsub ptr %ptr, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fsub_noret_v2bf16__offset12b_pos(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB54_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB54_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB54_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB54_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB54_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB54_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB54_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: .LBB54_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[4:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB54_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 511
+ %unused = atomicrmw fsub ptr %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define void @flat_agent_atomic_fsub_noret_v2bf16__offset12b_neg(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_agent_atomic_fsub_noret_v2bf16__offset12b_neg:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:-2048
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:-2048 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB55_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_agent_atomic_fsub_noret_v2bf16__offset12b_neg:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX940-NEXT: s_movk_i32 s0, 0xf800
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX940-NEXT: flat_load_dword v3, v[4:5]
+; GFX940-NEXT: s_mov_b32 s1, -1
+; GFX940-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[0:1]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB55_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_agent_atomic_fsub_noret_v2bf16__offset12b_neg:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX11-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX11-NEXT: flat_load_b32 v3, v[3:4]
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB55_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_agent_atomic_fsub_noret_v2bf16__offset12b_neg:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0xfffff800, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, -1, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB55_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_agent_atomic_fsub_noret_v2bf16__offset12b_neg:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_co_u32_e32 v4, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v5, vcc, -1, v1, vcc
+; GFX90A-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX90A-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX90A-NEXT: flat_load_dword v1, v[0:1]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX90A-NEXT: v_sub_f32_e32 v0, v0, v3
+; GFX90A-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX90A-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v0, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX90A-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v0, v6, v0, s9
+; GFX90A-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB55_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_agent_atomic_fsub_noret_v2bf16__offset12b_neg:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_co_u32_e32 v3, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v4, vcc, -1, v1, vcc
+; GFX908-NEXT: v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX908-NEXT: v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
+; GFX908-NEXT: flat_load_dword v1, v[0:1]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX908-NEXT: v_sub_f32_e32 v0, v0, v5
+; GFX908-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX908-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v0, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX908-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v0, v6, v0, s9
+; GFX908-NEXT: flat_atomic_cmpswap v0, v[3:4], v[0:1] glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB55_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_agent_atomic_fsub_noret_v2bf16__offset12b_neg:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0xfffff800, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB55_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_agent_atomic_fsub_noret_v2bf16__offset12b_neg:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0xfffff800, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, -1, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: .LBB55_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[4:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB55_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 -512
+ %unused = atomicrmw fsub ptr %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x bfloat> @flat_system_atomic_fsub_ret_v2bf16__offset12b_pos(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX12-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX12-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX12-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX12-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB56_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v3
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX940-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX940-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v3, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v5, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v6, v5, v3, s5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB56_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v3
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v3, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v3, v7, v9, s0
+; GFX11-NEXT: v_perm_b32 v5, v5, v3, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[5:6] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v6
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB56_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v4, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v0, v[3:4]
+; GFX10-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: v_sub_f32_e32 v0, v0, v1
+; GFX10-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX10-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v0
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v0, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v5, v5, v0, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB56_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v3
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v3, s9
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v3, v[0:1], v[6:7] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB56_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v3
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_sub_f32_e32 v3, v3, v4
+; GFX908-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX908-NEXT: v_bfe_u32 v7, v3, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v7, v7, v3, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v5, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3
+; GFX908-NEXT: v_cndmask_b32_e64 v3, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v3, s9
+; GFX908-NEXT: flat_atomic_cmpswap v3, v[0:1], v[5:6] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB56_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v0, v[3:4]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX8-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_sub_f32_e32 v0, v0, v1
+; GFX8-NEXT: v_sub_f32_e32 v5, v5, v2
+; GFX8-NEXT: v_bfe_u32 v7, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v0
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v0, 16
+; GFX8-NEXT: flat_atomic_cmpswap v0, v[3:4], v[5:6] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB56_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fsub_ret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v0, v[4:5]
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: .LBB56_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v0, v0, v6, 16
+; GFX7-NEXT: flat_atomic_cmpswap v0, v[4:5], v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB56_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 511
+ %result = atomicrmw fsub ptr %gep, <2 x bfloat> %val seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @flat_system_atomic_fsub_noret_v2bf16__offset12b_pos(ptr %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: flat_system_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX12-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX12-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX12-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX12-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX12-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX12-NEXT: v_mov_b32_e32 v3, v2
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB57_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: flat_system_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX940-NEXT: s_movk_i32 s4, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX940-NEXT: s_mov_b32 s5, 0x7060302
+; GFX940-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX940-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX940-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX940-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX940-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX940-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX940-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX940-NEXT: v_add3_u32 v7, v7, v2, s4
+; GFX940-NEXT: v_add3_u32 v9, v9, v6, s4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX940-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[0:1]
+; GFX940-NEXT: v_perm_b32 v2, v6, v2, s5
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 sc0 sc1
+; GFX940-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB57_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: flat_system_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: flat_load_b32 v3, v[0:1] offset:2044
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX11-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX11-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s0, v2, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v2, v7, v9, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2044 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX11-NEXT: v_mov_b32_e32 v3, v2
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB57_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: flat_system_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_co_u32 v0, vcc_lo, 0x7fc, v0
+; GFX10-NEXT: v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_load_dword v3, v[0:1]
+; GFX10-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX10-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX10-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX10-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v2
+; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX10-NEXT: v_add3_u32 v7, v7, v2, 0x7fff
+; GFX10-NEXT: v_add3_u32 v8, v8, v6, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s4, v2, v2
+; GFX10-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v2, v7, v9, s4
+; GFX10-NEXT: v_perm_b32 v2, v6, v2, 0x7060302
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3
+; GFX10-NEXT: v_mov_b32_e32 v3, v2
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB57_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: flat_system_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX90A-NEXT: s_movk_i32 s8, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX90A-NEXT: s_mov_b32 s9, 0x7060302
+; GFX90A-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX90A-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX90A-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX90A-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX90A-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX90A-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX90A-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX90A-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX90A-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX90A-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX90A-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB57_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: flat_system_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: flat_load_dword v3, v[0:1] offset:2044
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX908-NEXT: s_movk_i32 s8, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX908-NEXT: s_mov_b32 s9, 0x7060302
+; GFX908-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX908-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX908-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX908-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX908-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX908-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX908-NEXT: v_add3_u32 v7, v7, v2, s8
+; GFX908-NEXT: v_add3_u32 v9, v9, v6, s8
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX908-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX908-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX908-NEXT: v_perm_b32 v2, v6, v2, s9
+; GFX908-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:2044 glc
+; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v3, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB57_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: flat_system_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x7fc, v0
+; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT: flat_load_dword v3, v[0:1]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v2
+; GFX8-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
+; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX8-NEXT: v_sub_f32_e32 v2, v2, v4
+; GFX8-NEXT: v_sub_f32_e32 v6, v6, v5
+; GFX8-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v2
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
+; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v2
+; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2
+; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v2, v7, v8, s[4:5]
+; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
+; GFX8-NEXT: v_alignbit_b32 v2, v6, v2, 16
+; GFX8-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v3, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB57_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: flat_system_atomic_fsub_noret_v2bf16__offset12b_pos:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x7fc, v0
+; GFX7-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT: flat_load_dword v5, v[0:1]
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: s_mov_b64 s[4:5], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: .LBB57_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_sub_f32_e32 v7, v7, v3
+; GFX7-NEXT: v_sub_f32_e32 v6, v6, v2
+; GFX7-NEXT: v_alignbit_b32 v5, v4, v5, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: flat_atomic_cmpswap v6, v[0:1], v[4:5] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v6
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB57_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr %ptr, i64 511
+ %unused = atomicrmw fsub ptr %gep, <2 x bfloat> %val seq_cst
+ ret void
+}
+
+attributes #0 = { nounwind "amdgpu-unsafe-fp-atomics"="true" }
+attributes #1 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" "amdgpu-unsafe-fp-atomics"="true" }
>From 69249ebb7dce535314e22372238ed522cfa90caf Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Wed, 12 Jun 2024 12:54:17 +0200
Subject: [PATCH 4/4] AMDGPU: Add buffer atomicrmw tests for fadd/fmin/fmax
---
.../buffer-fat-pointer-atomicrmw-fadd.ll | 7394 +++++++++++++++
.../buffer-fat-pointer-atomicrmw-fmax.ll | 8062 +++++++++++++++++
.../buffer-fat-pointer-atomicrmw-fmin.ll | 8062 +++++++++++++++++
3 files changed, 23518 insertions(+)
create mode 100644 llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fadd.ll
create mode 100644 llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmax.ll
create mode 100644 llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmin.ll
diff --git a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fadd.ll b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fadd.ll
new file mode 100644
index 0000000000000..2f4606035376d
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fadd.ll
@@ -0,0 +1,7394 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 < %s | FileCheck -check-prefix=GFX12 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx940 < %s | FileCheck -check-prefix=GFX940 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx90a < %s | FileCheck -check-prefix=GFX90A %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx908 < %s | FileCheck -check-prefix=GFX908 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tonga < %s | FileCheck -check-prefix=GFX8 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=hawaii < %s | FileCheck -check-prefix=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tahiti < %s | FileCheck -check-prefix=GFX6 %s
+
+; --------------------------------------------------------------------
+; float
+; --------------------------------------------------------------------
+
+define float @buffer_fat_ptr_agent_atomic_fadd_ret_f32__offset(ptr addrspace(7) inreg %ptr, float %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f32__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v1, s4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: buffer_atomic_add_f32 v0, v1, s[0:3], null offen offset:1024 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f32__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, s4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: buffer_atomic_add_f32 v0, v1, s[0:3], 0 offen offset:1024 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f32__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v1, s4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: buffer_atomic_add_f32 v0, v1, s[0:3], 0 offen offset:1024 glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f32__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v0
+; GFX10-NEXT: v_mov_b32_e32 v0, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: v_mov_b32_e32 v3, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v5, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB0_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f32__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, s8
+; GFX90A-NEXT: buffer_atomic_add_f32 v0, v1, s[4:7], 0 offen offset:1024 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f32__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v0
+; GFX908-NEXT: v_mov_b32_e32 v0, s8
+; GFX908-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_add_i32 s10, s8, 0x400
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_mov_b32_e32 v3, s10
+; GFX908-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v5, v0
+; GFX908-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB0_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f32__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v0
+; GFX8-NEXT: v_mov_b32_e32 v0, s8
+; GFX8-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_add_i32 s10, s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_mov_b32_e32 v3, s10
+; GFX8-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v5, v0
+; GFX8-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: v_mov_b32_e32 v1, v5
+; GFX8-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB0_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f32__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v0
+; GFX7-NEXT: v_mov_b32_e32 v0, s8
+; GFX7-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: s_add_i32 s10, s8, 0x400
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mov_b32_e32 v3, s10
+; GFX7-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v0
+; GFX7-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v0, v4
+; GFX7-NEXT: v_mov_b32_e32 v1, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB0_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f32__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v0
+; GFX6-NEXT: v_mov_b32_e32 v0, s8
+; GFX6-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: s_add_i32 s10, s8, 0x400
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mov_b32_e32 v3, s10
+; GFX6-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v0
+; GFX6-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v0, v4
+; GFX6-NEXT: v_mov_b32_e32 v1, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB0_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fadd ptr addrspace(7) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define void @buffer_fat_ptr_agent_atomic_fadd_noret_f32__offset(ptr addrspace(7) inreg %ptr, float %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f32__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v1, s4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: buffer_atomic_add_f32 v0, v1, s[0:3], null offen offset:1024
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f32__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, s4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: buffer_atomic_add_f32 v0, v1, s[0:3], 0 offen offset:1024
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f32__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v1, s4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: buffer_atomic_add_f32 v0, v1, s[0:3], 0 offen offset:1024
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f32__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: v_mov_b32_e32 v3, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v2, v1, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f32_e32 v1, v2, v0
+; GFX10-NEXT: v_mov_b32_e32 v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_mov_b32_e32 v4, v1
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX10-NEXT: v_mov_b32_e32 v2, v4
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB1_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f32__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, s8
+; GFX90A-NEXT: buffer_atomic_add_f32 v0, v1, s[4:7], 0 offen offset:1024
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f32__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, s8
+; GFX908-NEXT: buffer_atomic_add_f32 v0, v1, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f32__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, s8
+; GFX8-NEXT: buffer_load_dword v2, v1, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_add_i32 s10, s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_mov_b32_e32 v3, s10
+; GFX8-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f32_e32 v1, v2, v0
+; GFX8-NEXT: v_mov_b32_e32 v5, v2
+; GFX8-NEXT: v_mov_b32_e32 v4, v1
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v2
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v2, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB1_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f32__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v1, s8
+; GFX7-NEXT: buffer_load_dword v2, v1, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: s_add_i32 s10, s8, 0x400
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mov_b32_e32 v3, s10
+; GFX7-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v1, v2, v0
+; GFX7-NEXT: v_mov_b32_e32 v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v4, v1
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v2
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v2, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB1_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f32__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v1, s8
+; GFX6-NEXT: buffer_load_dword v2, v1, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: s_add_i32 s10, s8, 0x400
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mov_b32_e32 v3, s10
+; GFX6-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v1, v2, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v2
+; GFX6-NEXT: v_mov_b32_e32 v4, v1
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v2
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v2, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB1_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(7) %ptr, i32 256
+ %unused = atomicrmw fadd ptr addrspace(7) %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define float @buffer_fat_ptr_agent_atomic_fadd_ret_f32__offset__waterfall(ptr addrspace(7) %ptr, float %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f32__offset__waterfall:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_mov_b32 s1, exec_lo
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: buffer_atomic_add_f32 v5, v4, s[4:7], null offen offset:1024 th:TH_ATOMIC_RETURN
+; GFX12-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
+; GFX12-NEXT: ; implicit-def: $vgpr4
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB2_1
+; GFX12-NEXT: ; %bb.2:
+; GFX12-NEXT: s_mov_b32 exec_lo, s1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v0, v5
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f32__offset__waterfall:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_mov_b64 s[2:3], exec
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_atomic_add_f32 v5, v4, s[4:7], 0 offen offset:1024 sc0
+; GFX940-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
+; GFX940-NEXT: ; implicit-def: $vgpr4
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB2_1
+; GFX940-NEXT: ; %bb.2:
+; GFX940-NEXT: s_mov_b64 exec, s[2:3]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v0, v5
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f32__offset__waterfall:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_mov_b32 s1, exec_lo
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_atomic_add_f32 v5, v4, s[4:7], 0 offen offset:1024 glc
+; GFX11-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
+; GFX11-NEXT: ; implicit-def: $vgpr4
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB2_1
+; GFX11-NEXT: ; %bb.2:
+; GFX11-NEXT: s_mov_b32 exec_lo, s1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v0, v5
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f32__offset__waterfall:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v9, 0x400, v4
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: buffer_load_dword v8, v4, s[8:11], 0 offen offset:1024
+; GFX10-NEXT: ; implicit-def: $vgpr4
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB2_1
+; GFX10-NEXT: ; %bb.2:
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Loop Header: Depth=1
+; GFX10-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f32_e32 v7, v8, v5
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_mov_b32_e32 v6, v7
+; GFX10-NEXT: v_mov_b32_e32 v7, v8
+; GFX10-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX10-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_atomic_cmpswap v[6:7], v9, s[8:11], 0 offen glc
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB2_4
+; GFX10-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v6, v8
+; GFX10-NEXT: v_mov_b32_e32 v8, v6
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB2_3
+; GFX10-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v6
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f32__offset__waterfall:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: s_mov_b64 s[6:7], exec
+; GFX90A-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_atomic_add_f32 v5, v4, s[8:11], 0 offen offset:1024 glc
+; GFX90A-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
+; GFX90A-NEXT: ; implicit-def: $vgpr4
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB2_1
+; GFX90A-NEXT: ; %bb.2:
+; GFX90A-NEXT: s_mov_b64 exec, s[6:7]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v0, v5
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f32__offset__waterfall:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v9, 0x400, v4
+; GFX908-NEXT: s_mov_b64 s[6:7], exec
+; GFX908-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_nop 0
+; GFX908-NEXT: buffer_load_dword v8, v4, s[8:11], 0 offen offset:1024
+; GFX908-NEXT: ; implicit-def: $vgpr4
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB2_1
+; GFX908-NEXT: ; %bb.2:
+; GFX908-NEXT: s_mov_b64 exec, s[6:7]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Loop Header: Depth=1
+; GFX908-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f32_e32 v7, v8, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v7
+; GFX908-NEXT: s_mov_b64 s[12:13], exec
+; GFX908-NEXT: v_mov_b32_e32 v7, v8
+; GFX908-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX908-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_atomic_cmpswap v[6:7], v9, s[8:11], 0 offen glc
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB2_4
+; GFX908-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX908-NEXT: s_mov_b64 exec, s[12:13]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v6, v8
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v8, v6
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB2_3
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v6
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f32__offset__waterfall:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x400, v4
+; GFX8-NEXT: s_mov_b64 s[6:7], exec
+; GFX8-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_nop 0
+; GFX8-NEXT: buffer_load_dword v8, v4, s[8:11], 0 offen offset:1024
+; GFX8-NEXT: ; implicit-def: $vgpr4
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB2_1
+; GFX8-NEXT: ; %bb.2:
+; GFX8-NEXT: s_mov_b64 exec, s[6:7]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Loop Header: Depth=1
+; GFX8-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f32_e32 v7, v8, v5
+; GFX8-NEXT: v_mov_b32_e32 v6, v7
+; GFX8-NEXT: s_mov_b64 s[12:13], exec
+; GFX8-NEXT: v_mov_b32_e32 v7, v8
+; GFX8-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX8-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_atomic_cmpswap v[6:7], v9, s[8:11], 0 offen glc
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB2_4
+; GFX8-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX8-NEXT: s_mov_b64 exec, s[12:13]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v6, v8
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v8, v6
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB2_3
+; GFX8-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v6
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f32__offset__waterfall:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v9, vcc, 0x400, v4
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: buffer_load_dword v8, v4, s[8:11], 0 offen offset:1024
+; GFX7-NEXT: ; implicit-def: $vgpr4
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_1
+; GFX7-NEXT: ; %bb.2:
+; GFX7-NEXT: s_mov_b64 exec, s[6:7]
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Loop Header: Depth=1
+; GFX7-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_f32_e32 v7, v8, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v7
+; GFX7-NEXT: s_mov_b64 s[12:13], exec
+; GFX7-NEXT: v_mov_b32_e32 v7, v8
+; GFX7-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX7-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v9, s[8:11], 0 offen glc
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_4
+; GFX7-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX7-NEXT: s_mov_b64 exec, s[12:13]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v8
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_3
+; GFX7-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v0, v6
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f32__offset__waterfall:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v9, vcc, 0x400, v4
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: buffer_load_dword v8, v4, s[8:11], 0 offen offset:1024
+; GFX6-NEXT: ; implicit-def: $vgpr4
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB2_1
+; GFX6-NEXT: ; %bb.2:
+; GFX6-NEXT: s_mov_b64 exec, s[6:7]
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Loop Header: Depth=1
+; GFX6-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: v_add_f32_e32 v7, v8, v5
+; GFX6-NEXT: v_mov_b32_e32 v6, v7
+; GFX6-NEXT: s_mov_b64 s[12:13], exec
+; GFX6-NEXT: v_mov_b32_e32 v7, v8
+; GFX6-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX6-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v9, s[8:11], 0 offen glc
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB2_4
+; GFX6-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX6-NEXT: s_mov_b64 exec, s[12:13]
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v8
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB2_3
+; GFX6-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v0, v6
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fadd ptr addrspace(7) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+; --------------------------------------------------------------------
+; double
+; --------------------------------------------------------------------
+
+define double @buffer_fat_ptr_agent_atomic_fadd_ret_f64__offset(ptr addrspace(7) inreg %ptr, double %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f64__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
+; GFX12-NEXT: v_mov_b32_e32 v0, s4
+; GFX12-NEXT: s_addk_co_i32 s4, 0x800
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_mov_b32_e32 v6, s4
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: buffer_load_b64 v[0:1], v0, s[0:3], null offen offset:2048
+; GFX12-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v10, v1 :: v_dual_mov_b32 v9, v0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f64_e32 v[7:8], v[9:10], v[4:5]
+; GFX12-NEXT: v_dual_mov_b32 v0, v7 :: v_dual_mov_b32 v1, v8
+; GFX12-NEXT: v_dual_mov_b32 v2, v9 :: v_dual_mov_b32 v3, v10
+; GFX12-NEXT: buffer_atomic_cmpswap_b64 v[0:3], v6, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[9:10]
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB3_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f64__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, s4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: buffer_atomic_add_f64 v[0:1], v2, s[0:3], 0 offen offset:2048 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f64__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
+; GFX11-NEXT: v_mov_b32_e32 v0, s4
+; GFX11-NEXT: s_addk_i32 s4, 0x800
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v6, s4
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: buffer_load_b64 v[0:1], v0, s[0:3], 0 offen offset:2048
+; GFX11-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v10, v1 :: v_dual_mov_b32 v9, v0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f64 v[7:8], v[9:10], v[4:5]
+; GFX11-NEXT: v_dual_mov_b32 v0, v7 :: v_dual_mov_b32 v1, v8
+; GFX11-NEXT: v_dual_mov_b32 v2, v9 :: v_dual_mov_b32 v3, v10
+; GFX11-NEXT: buffer_atomic_cmpswap_b64 v[0:3], v6, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[9:10]
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB3_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f64__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v4, v0
+; GFX10-NEXT: v_mov_b32_e32 v0, s8
+; GFX10-NEXT: v_mov_b32_e32 v5, v1
+; GFX10-NEXT: s_addk_i32 s8, 0x800
+; GFX10-NEXT: v_mov_b32_e32 v6, s8
+; GFX10-NEXT: buffer_load_dwordx2 v[0:1], v0, s[4:7], 0 offen offset:2048
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v10, v1
+; GFX10-NEXT: v_mov_b32_e32 v9, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_add_f64 v[7:8], v[9:10], v[4:5]
+; GFX10-NEXT: v_mov_b32_e32 v0, v7
+; GFX10-NEXT: v_mov_b32_e32 v1, v8
+; GFX10-NEXT: v_mov_b32_e32 v2, v9
+; GFX10-NEXT: v_mov_b32_e32 v3, v10
+; GFX10-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v6, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[9:10]
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB3_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f64__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, s8
+; GFX90A-NEXT: buffer_atomic_add_f64 v[0:1], v2, s[4:7], 0 offen offset:2048 glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f64__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v4, v0
+; GFX908-NEXT: v_mov_b32_e32 v0, s8
+; GFX908-NEXT: v_mov_b32_e32 v5, v1
+; GFX908-NEXT: buffer_load_dwordx2 v[0:1], v0, s[4:7], 0 offen offset:2048
+; GFX908-NEXT: s_add_i32 s10, s8, 0x800
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_mov_b32_e32 v6, s10
+; GFX908-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v10, v1
+; GFX908-NEXT: v_mov_b32_e32 v9, v0
+; GFX908-NEXT: v_add_f64 v[7:8], v[9:10], v[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v7
+; GFX908-NEXT: v_mov_b32_e32 v1, v8
+; GFX908-NEXT: v_mov_b32_e32 v2, v9
+; GFX908-NEXT: v_mov_b32_e32 v3, v10
+; GFX908-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v6, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[9:10]
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB3_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f64__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v4, v0
+; GFX8-NEXT: v_mov_b32_e32 v0, s8
+; GFX8-NEXT: v_mov_b32_e32 v5, v1
+; GFX8-NEXT: buffer_load_dwordx2 v[0:1], v0, s[4:7], 0 offen offset:2048
+; GFX8-NEXT: s_add_i32 s10, s8, 0x800
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_mov_b32_e32 v6, s10
+; GFX8-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v10, v1
+; GFX8-NEXT: v_mov_b32_e32 v9, v0
+; GFX8-NEXT: v_add_f64 v[7:8], v[9:10], v[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v7
+; GFX8-NEXT: v_mov_b32_e32 v1, v8
+; GFX8-NEXT: v_mov_b32_e32 v2, v9
+; GFX8-NEXT: v_mov_b32_e32 v3, v10
+; GFX8-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v6, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[9:10]
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB3_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f64__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v4, v0
+; GFX7-NEXT: v_mov_b32_e32 v0, s8
+; GFX7-NEXT: v_mov_b32_e32 v5, v1
+; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v0, s[4:7], 0 offen offset:2048
+; GFX7-NEXT: s_add_i32 s10, s8, 0x800
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mov_b32_e32 v6, s10
+; GFX7-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v10, v1
+; GFX7-NEXT: v_mov_b32_e32 v9, v0
+; GFX7-NEXT: v_add_f64 v[7:8], v[9:10], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v7
+; GFX7-NEXT: v_mov_b32_e32 v1, v8
+; GFX7-NEXT: v_mov_b32_e32 v2, v9
+; GFX7-NEXT: v_mov_b32_e32 v3, v10
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v6, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[9:10]
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB3_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f64__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v4, v0
+; GFX6-NEXT: v_mov_b32_e32 v0, s8
+; GFX6-NEXT: v_mov_b32_e32 v5, v1
+; GFX6-NEXT: buffer_load_dwordx2 v[0:1], v0, s[4:7], 0 offen offset:2048
+; GFX6-NEXT: s_add_i32 s10, s8, 0x800
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mov_b32_e32 v6, s10
+; GFX6-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v10, v1
+; GFX6-NEXT: v_mov_b32_e32 v9, v0
+; GFX6-NEXT: v_add_f64 v[7:8], v[9:10], v[4:5]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v0, v7
+; GFX6-NEXT: v_mov_b32_e32 v1, v8
+; GFX6-NEXT: v_mov_b32_e32 v2, v9
+; GFX6-NEXT: v_mov_b32_e32 v3, v10
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v6, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[9:10]
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB3_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fadd ptr addrspace(7) %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define void @buffer_fat_ptr_agent_atomic_fadd_noret_f64__offset(ptr addrspace(7) inreg %ptr, double %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f64__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, s4
+; GFX12-NEXT: s_addk_co_i32 s4, 0x800
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_mov_b32_e32 v6, s4
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: buffer_load_b64 v[4:5], v2, s[0:3], null offen offset:2048
+; GFX12-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_add_f64_e32 v[2:3], v[4:5], v[0:1]
+; GFX12-NEXT: v_dual_mov_b32 v10, v5 :: v_dual_mov_b32 v9, v4
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_dual_mov_b32 v8, v3 :: v_dual_mov_b32 v7, v2
+; GFX12-NEXT: buffer_atomic_cmpswap_b64 v[7:10], v6, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[7:8], v[4:5]
+; GFX12-NEXT: v_dual_mov_b32 v4, v7 :: v_dual_mov_b32 v5, v8
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB4_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f64__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, s4
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: buffer_atomic_add_f64 v[0:1], v2, s[0:3], 0 offen offset:2048
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f64__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, s4
+; GFX11-NEXT: s_addk_i32 s4, 0x800
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v6, s4
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: buffer_load_b64 v[4:5], v2, s[0:3], 0 offen offset:2048
+; GFX11-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_add_f64 v[2:3], v[4:5], v[0:1]
+; GFX11-NEXT: v_dual_mov_b32 v10, v5 :: v_dual_mov_b32 v9, v4
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_dual_mov_b32 v8, v3 :: v_dual_mov_b32 v7, v2
+; GFX11-NEXT: buffer_atomic_cmpswap_b64 v[7:10], v6, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[7:8], v[4:5]
+; GFX11-NEXT: v_dual_mov_b32 v4, v7 :: v_dual_mov_b32 v5, v8
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB4_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f64__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x800
+; GFX10-NEXT: v_mov_b32_e32 v6, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dwordx2 v[4:5], v2, s[4:7], 0 offen offset:2048
+; GFX10-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f64 v[2:3], v[4:5], v[0:1]
+; GFX10-NEXT: v_mov_b32_e32 v10, v5
+; GFX10-NEXT: v_mov_b32_e32 v9, v4
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_mov_b32_e32 v8, v3
+; GFX10-NEXT: v_mov_b32_e32 v7, v2
+; GFX10-NEXT: buffer_atomic_cmpswap_x2 v[7:10], v6, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[7:8], v[4:5]
+; GFX10-NEXT: v_mov_b32_e32 v4, v7
+; GFX10-NEXT: v_mov_b32_e32 v5, v8
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB4_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f64__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, s8
+; GFX90A-NEXT: buffer_atomic_add_f64 v[0:1], v2, s[4:7], 0 offen offset:2048
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f64__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, s8
+; GFX908-NEXT: buffer_load_dwordx2 v[4:5], v2, s[4:7], 0 offen offset:2048
+; GFX908-NEXT: s_add_i32 s10, s8, 0x800
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_mov_b32_e32 v6, s10
+; GFX908-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f64 v[2:3], v[4:5], v[0:1]
+; GFX908-NEXT: v_mov_b32_e32 v10, v5
+; GFX908-NEXT: v_mov_b32_e32 v9, v4
+; GFX908-NEXT: v_mov_b32_e32 v8, v3
+; GFX908-NEXT: v_mov_b32_e32 v7, v2
+; GFX908-NEXT: buffer_atomic_cmpswap_x2 v[7:10], v6, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[7:8], v[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v4, v7
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: v_mov_b32_e32 v5, v8
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB4_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f64__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, s8
+; GFX8-NEXT: buffer_load_dwordx2 v[4:5], v2, s[4:7], 0 offen offset:2048
+; GFX8-NEXT: s_add_i32 s10, s8, 0x800
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_mov_b32_e32 v6, s10
+; GFX8-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f64 v[2:3], v[4:5], v[0:1]
+; GFX8-NEXT: v_mov_b32_e32 v10, v5
+; GFX8-NEXT: v_mov_b32_e32 v9, v4
+; GFX8-NEXT: v_mov_b32_e32 v8, v3
+; GFX8-NEXT: v_mov_b32_e32 v7, v2
+; GFX8-NEXT: buffer_atomic_cmpswap_x2 v[7:10], v6, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[7:8], v[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v4, v7
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v5, v8
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB4_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f64__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dwordx2 v[4:5], v2, s[4:7], 0 offen offset:2048
+; GFX7-NEXT: s_add_i32 s10, s8, 0x800
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mov_b32_e32 v6, s10
+; GFX7-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_f64 v[2:3], v[4:5], v[0:1]
+; GFX7-NEXT: v_mov_b32_e32 v10, v5
+; GFX7-NEXT: v_mov_b32_e32 v9, v4
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[7:10], v6, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[7:8], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v4, v7
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v5, v8
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB4_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f64__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dwordx2 v[4:5], v2, s[4:7], 0 offen offset:2048
+; GFX6-NEXT: s_add_i32 s10, s8, 0x800
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mov_b32_e32 v6, s10
+; GFX6-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_add_f64 v[2:3], v[4:5], v[0:1]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v10, v5
+; GFX6-NEXT: v_mov_b32_e32 v9, v4
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[7:10], v6, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[7:8], v[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v4, v7
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v5, v8
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB4_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(7) %ptr, i32 256
+ %unused = atomicrmw fadd ptr addrspace(7) %gep, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+define double @buffer_fat_ptr_agent_atomic_fadd_ret_f64__offset__waterfall(ptr addrspace(7) %ptr, double %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f64__offset__waterfall:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v8, v3 :: v_dual_mov_b32 v7, v2
+; GFX12-NEXT: v_dual_mov_b32 v10, v1 :: v_dual_mov_b32 v9, v0
+; GFX12-NEXT: v_add_nc_u32_e32 v15, 0x800, v4
+; GFX12-NEXT: s_mov_b32 s1, exec_lo
+; GFX12-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_readfirstlane_b32 s4, v9
+; GFX12-NEXT: v_readfirstlane_b32 s5, v10
+; GFX12-NEXT: v_readfirstlane_b32 s6, v7
+; GFX12-NEXT: v_readfirstlane_b32 s7, v8
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[9:10]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[7:8]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: buffer_load_b64 v[13:14], v4, s[4:7], null offen offset:2048
+; GFX12-NEXT: ; implicit-def: $vgpr4
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB5_1
+; GFX12-NEXT: ; %bb.2:
+; GFX12-NEXT: s_mov_b32 exec_lo, s1
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Loop Header: Depth=1
+; GFX12-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_add_f64_e32 v[11:12], v[13:14], v[5:6]
+; GFX12-NEXT: s_mov_b32 s2, exec_lo
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v0, v11 :: v_dual_mov_b32 v1, v12
+; GFX12-NEXT: v_dual_mov_b32 v2, v13 :: v_dual_mov_b32 v3, v14
+; GFX12-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX12-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX12-NEXT: v_readfirstlane_b32 s4, v9
+; GFX12-NEXT: v_readfirstlane_b32 s5, v10
+; GFX12-NEXT: v_readfirstlane_b32 s6, v7
+; GFX12-NEXT: v_readfirstlane_b32 s7, v8
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[9:10]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[7:8]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: buffer_atomic_cmpswap_b64 v[0:3], v15, s[4:7], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB5_4
+; GFX12-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX12-NEXT: s_mov_b32 exec_lo, s2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[13:14]
+; GFX12-NEXT: v_dual_mov_b32 v14, v1 :: v_dual_mov_b32 v13, v0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB5_3
+; GFX12-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f64__offset__waterfall:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v6
+; GFX940-NEXT: v_mov_b32_e32 v6, v5
+; GFX940-NEXT: s_mov_b64 s[2:3], exec
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_atomic_add_f64 v[6:7], v4, s[4:7], 0 offen offset:2048 sc0
+; GFX940-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
+; GFX940-NEXT: ; implicit-def: $vgpr4
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB5_1
+; GFX940-NEXT: ; %bb.2:
+; GFX940-NEXT: s_mov_b64 exec, s[2:3]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v0, v6
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f64__offset__waterfall:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v8, v3 :: v_dual_mov_b32 v7, v2
+; GFX11-NEXT: v_dual_mov_b32 v10, v1 :: v_dual_mov_b32 v9, v0
+; GFX11-NEXT: v_add_nc_u32_e32 v15, 0x800, v4
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_readfirstlane_b32 s4, v9
+; GFX11-NEXT: v_readfirstlane_b32 s5, v10
+; GFX11-NEXT: v_readfirstlane_b32 s6, v7
+; GFX11-NEXT: v_readfirstlane_b32 s7, v8
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[9:10]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[7:8]
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: buffer_load_b64 v[13:14], v4, s[4:7], 0 offen offset:2048
+; GFX11-NEXT: ; implicit-def: $vgpr4
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB5_1
+; GFX11-NEXT: ; %bb.2:
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Loop Header: Depth=1
+; GFX11-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_add_f64 v[11:12], v[13:14], v[5:6]
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v0, v11 :: v_dual_mov_b32 v1, v12
+; GFX11-NEXT: v_dual_mov_b32 v2, v13 :: v_dual_mov_b32 v3, v14
+; GFX11-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX11-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX11-NEXT: v_readfirstlane_b32 s4, v9
+; GFX11-NEXT: v_readfirstlane_b32 s5, v10
+; GFX11-NEXT: v_readfirstlane_b32 s6, v7
+; GFX11-NEXT: v_readfirstlane_b32 s7, v8
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[9:10]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[7:8]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_atomic_cmpswap_b64 v[0:3], v15, s[4:7], 0 offen glc
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB5_4
+; GFX11-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[13:14]
+; GFX11-NEXT: v_dual_mov_b32 v14, v1 :: v_dual_mov_b32 v13, v0
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB5_3
+; GFX11-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f64__offset__waterfall:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v8, v3
+; GFX10-NEXT: v_mov_b32_e32 v7, v2
+; GFX10-NEXT: v_mov_b32_e32 v10, v1
+; GFX10-NEXT: v_mov_b32_e32 v9, v0
+; GFX10-NEXT: v_add_nc_u32_e32 v15, 0x800, v4
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_readfirstlane_b32 s8, v9
+; GFX10-NEXT: v_readfirstlane_b32 s9, v10
+; GFX10-NEXT: v_readfirstlane_b32 s10, v7
+; GFX10-NEXT: v_readfirstlane_b32 s11, v8
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[9:10]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[7:8]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: buffer_load_dwordx2 v[13:14], v4, s[8:11], 0 offen offset:2048
+; GFX10-NEXT: ; implicit-def: $vgpr4
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB5_1
+; GFX10-NEXT: ; %bb.2:
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Loop Header: Depth=1
+; GFX10-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f64 v[11:12], v[13:14], v[5:6]
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_mov_b32_e32 v0, v11
+; GFX10-NEXT: v_mov_b32_e32 v1, v12
+; GFX10-NEXT: v_mov_b32_e32 v2, v13
+; GFX10-NEXT: v_mov_b32_e32 v3, v14
+; GFX10-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX10-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX10-NEXT: v_readfirstlane_b32 s8, v9
+; GFX10-NEXT: v_readfirstlane_b32 s9, v10
+; GFX10-NEXT: v_readfirstlane_b32 s10, v7
+; GFX10-NEXT: v_readfirstlane_b32 s11, v8
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[9:10]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[7:8]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v15, s[8:11], 0 offen glc
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB5_4
+; GFX10-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[13:14]
+; GFX10-NEXT: v_mov_b32_e32 v14, v1
+; GFX10-NEXT: v_mov_b32_e32 v13, v0
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB5_3
+; GFX10-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f64__offset__waterfall:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v6
+; GFX90A-NEXT: v_mov_b32_e32 v6, v5
+; GFX90A-NEXT: s_mov_b64 s[6:7], exec
+; GFX90A-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_atomic_add_f64 v[6:7], v4, s[8:11], 0 offen offset:2048 glc
+; GFX90A-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
+; GFX90A-NEXT: ; implicit-def: $vgpr4
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB5_1
+; GFX90A-NEXT: ; %bb.2:
+; GFX90A-NEXT: s_mov_b64 exec, s[6:7]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v0, v6
+; GFX90A-NEXT: v_mov_b32_e32 v1, v7
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f64__offset__waterfall:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v8, v3
+; GFX908-NEXT: v_mov_b32_e32 v7, v2
+; GFX908-NEXT: v_mov_b32_e32 v10, v1
+; GFX908-NEXT: v_mov_b32_e32 v9, v0
+; GFX908-NEXT: v_add_u32_e32 v15, 0x800, v4
+; GFX908-NEXT: s_mov_b64 s[6:7], exec
+; GFX908-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: v_readfirstlane_b32 s8, v9
+; GFX908-NEXT: v_readfirstlane_b32 s9, v10
+; GFX908-NEXT: v_readfirstlane_b32 s10, v7
+; GFX908-NEXT: v_readfirstlane_b32 s11, v8
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_nop 0
+; GFX908-NEXT: buffer_load_dwordx2 v[13:14], v4, s[8:11], 0 offen offset:2048
+; GFX908-NEXT: ; implicit-def: $vgpr4
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_1
+; GFX908-NEXT: ; %bb.2:
+; GFX908-NEXT: s_mov_b64 exec, s[6:7]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Loop Header: Depth=1
+; GFX908-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f64 v[11:12], v[13:14], v[5:6]
+; GFX908-NEXT: s_mov_b64 s[12:13], exec
+; GFX908-NEXT: v_mov_b32_e32 v0, v11
+; GFX908-NEXT: v_mov_b32_e32 v1, v12
+; GFX908-NEXT: v_mov_b32_e32 v2, v13
+; GFX908-NEXT: v_mov_b32_e32 v3, v14
+; GFX908-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX908-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX908-NEXT: v_readfirstlane_b32 s8, v9
+; GFX908-NEXT: v_readfirstlane_b32 s9, v10
+; GFX908-NEXT: v_readfirstlane_b32 s10, v7
+; GFX908-NEXT: v_readfirstlane_b32 s11, v8
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v15, s[8:11], 0 offen glc
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_4
+; GFX908-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX908-NEXT: s_mov_b64 exec, s[12:13]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[13:14]
+; GFX908-NEXT: v_mov_b32_e32 v14, v1
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v13, v0
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_3
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f64__offset__waterfall:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v8, v3
+; GFX8-NEXT: v_mov_b32_e32 v7, v2
+; GFX8-NEXT: v_mov_b32_e32 v10, v1
+; GFX8-NEXT: v_mov_b32_e32 v9, v0
+; GFX8-NEXT: v_add_u32_e32 v15, vcc, 0x800, v4
+; GFX8-NEXT: s_mov_b64 s[6:7], exec
+; GFX8-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: v_readfirstlane_b32 s8, v9
+; GFX8-NEXT: v_readfirstlane_b32 s9, v10
+; GFX8-NEXT: v_readfirstlane_b32 s10, v7
+; GFX8-NEXT: v_readfirstlane_b32 s11, v8
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_nop 0
+; GFX8-NEXT: buffer_load_dwordx2 v[13:14], v4, s[8:11], 0 offen offset:2048
+; GFX8-NEXT: ; implicit-def: $vgpr4
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_1
+; GFX8-NEXT: ; %bb.2:
+; GFX8-NEXT: s_mov_b64 exec, s[6:7]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Loop Header: Depth=1
+; GFX8-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f64 v[11:12], v[13:14], v[5:6]
+; GFX8-NEXT: s_mov_b64 s[12:13], exec
+; GFX8-NEXT: v_mov_b32_e32 v0, v11
+; GFX8-NEXT: v_mov_b32_e32 v1, v12
+; GFX8-NEXT: v_mov_b32_e32 v2, v13
+; GFX8-NEXT: v_mov_b32_e32 v3, v14
+; GFX8-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX8-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX8-NEXT: v_readfirstlane_b32 s8, v9
+; GFX8-NEXT: v_readfirstlane_b32 s9, v10
+; GFX8-NEXT: v_readfirstlane_b32 s10, v7
+; GFX8-NEXT: v_readfirstlane_b32 s11, v8
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v15, s[8:11], 0 offen glc
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_4
+; GFX8-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX8-NEXT: s_mov_b64 exec, s[12:13]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[13:14]
+; GFX8-NEXT: v_mov_b32_e32 v14, v1
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v13, v0
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_3
+; GFX8-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f64__offset__waterfall:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v10, v1
+; GFX7-NEXT: v_mov_b32_e32 v9, v0
+; GFX7-NEXT: v_add_i32_e32 v15, vcc, 0x800, v4
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_readfirstlane_b32 s8, v9
+; GFX7-NEXT: v_readfirstlane_b32 s9, v10
+; GFX7-NEXT: v_readfirstlane_b32 s10, v7
+; GFX7-NEXT: v_readfirstlane_b32 s11, v8
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: buffer_load_dwordx2 v[13:14], v4, s[8:11], 0 offen offset:2048
+; GFX7-NEXT: ; implicit-def: $vgpr4
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_1
+; GFX7-NEXT: ; %bb.2:
+; GFX7-NEXT: s_mov_b64 exec, s[6:7]
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Loop Header: Depth=1
+; GFX7-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_add_f64 v[11:12], v[13:14], v[5:6]
+; GFX7-NEXT: s_mov_b64 s[12:13], exec
+; GFX7-NEXT: v_mov_b32_e32 v0, v11
+; GFX7-NEXT: v_mov_b32_e32 v1, v12
+; GFX7-NEXT: v_mov_b32_e32 v2, v13
+; GFX7-NEXT: v_mov_b32_e32 v3, v14
+; GFX7-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX7-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX7-NEXT: v_readfirstlane_b32 s8, v9
+; GFX7-NEXT: v_readfirstlane_b32 s9, v10
+; GFX7-NEXT: v_readfirstlane_b32 s10, v7
+; GFX7-NEXT: v_readfirstlane_b32 s11, v8
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v15, s[8:11], 0 offen glc
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_4
+; GFX7-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX7-NEXT: s_mov_b64 exec, s[12:13]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[13:14]
+; GFX7-NEXT: v_mov_b32_e32 v14, v1
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v13, v0
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_3
+; GFX7-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f64__offset__waterfall:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v10, v1
+; GFX6-NEXT: v_mov_b32_e32 v9, v0
+; GFX6-NEXT: v_add_i32_e32 v15, vcc, 0x800, v4
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_readfirstlane_b32 s8, v9
+; GFX6-NEXT: v_readfirstlane_b32 s9, v10
+; GFX6-NEXT: v_readfirstlane_b32 s10, v7
+; GFX6-NEXT: v_readfirstlane_b32 s11, v8
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: buffer_load_dwordx2 v[13:14], v4, s[8:11], 0 offen offset:2048
+; GFX6-NEXT: ; implicit-def: $vgpr4
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB5_1
+; GFX6-NEXT: ; %bb.2:
+; GFX6-NEXT: s_mov_b64 exec, s[6:7]
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Loop Header: Depth=1
+; GFX6-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_add_f64 v[11:12], v[13:14], v[5:6]
+; GFX6-NEXT: s_mov_b64 s[12:13], exec
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v0, v11
+; GFX6-NEXT: v_mov_b32_e32 v1, v12
+; GFX6-NEXT: v_mov_b32_e32 v2, v13
+; GFX6-NEXT: v_mov_b32_e32 v3, v14
+; GFX6-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX6-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX6-NEXT: v_readfirstlane_b32 s8, v9
+; GFX6-NEXT: v_readfirstlane_b32 s9, v10
+; GFX6-NEXT: v_readfirstlane_b32 s10, v7
+; GFX6-NEXT: v_readfirstlane_b32 s11, v8
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v15, s[8:11], 0 offen glc
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB5_4
+; GFX6-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX6-NEXT: s_mov_b64 exec, s[12:13]
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[13:14]
+; GFX6-NEXT: v_mov_b32_e32 v14, v1
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v13, v0
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB5_3
+; GFX6-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fadd ptr addrspace(7) %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+; --------------------------------------------------------------------
+; half
+; --------------------------------------------------------------------
+
+define half @buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset(ptr addrspace(7) inreg %ptr, half %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_addk_co_i32 s4, 0x200
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s4, s4, -4
+; GFX12-NEXT: v_mov_b32_e32 v5, s4
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: buffer_load_b32 v2, v5, s[0:3], null offen
+; GFX12-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v1, 24, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f16_e32 v1, v1, v0
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 24, v1
+; GFX12-NEXT: v_and_or_b32 v1, 0xffffff, v2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v4, v2 :: v_dual_mov_b32 v3, v1
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[3:4], v5, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX12-NEXT: v_mov_b32_e32 v2, v3
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB6_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v3
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_addk_i32 s4, 0x200
+; GFX940-NEXT: s_and_b32 s4, s4, -4
+; GFX940-NEXT: v_mov_b32_e32 v1, s4
+; GFX940-NEXT: buffer_load_dword v3, v1, s[0:3], 0 offen
+; GFX940-NEXT: s_mov_b32 s6, 0xffffff
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_add_f16_sdwa v2, v3, v0 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:DWORD
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_and_or_b32 v2, v3, s6, v2
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[2:3]
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v1, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB6_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_addk_i32 s4, 0x200
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s4, s4, -4
+; GFX11-NEXT: v_mov_b32_e32 v5, s4
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: buffer_load_b32 v2, v5, s[0:3], 0 offen
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v1, 24, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f16_e32 v1, v1, v0
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 24, v1
+; GFX11-NEXT: v_and_or_b32 v1, 0xffffff, v2, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v4, v2 :: v_dual_mov_b32 v3, v1
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[3:4], v5, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX11-NEXT: v_mov_b32_e32 v2, v3
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB6_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v3
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_addk_i32 s8, 0x200
+; GFX10-NEXT: s_and_b32 s8, s8, -4
+; GFX10-NEXT: v_mov_b32_e32 v5, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v2, v5, s[4:7], 0 offen
+; GFX10-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f16_sdwa v1, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:DWORD
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 24, v1
+; GFX10-NEXT: v_and_or_b32 v1, 0xffffff, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v4, v2
+; GFX10-NEXT: v_mov_b32_e32 v3, v1
+; GFX10-NEXT: buffer_atomic_cmpswap v[3:4], v5, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2
+; GFX10-NEXT: v_mov_b32_e32 v2, v3
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB6_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 24, v3
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: s_addk_i32 s8, 0x200
+; GFX90A-NEXT: s_and_b32 s8, s8, -4
+; GFX90A-NEXT: v_mov_b32_e32 v1, s8
+; GFX90A-NEXT: buffer_load_dword v3, v1, s[4:7], 0 offen
+; GFX90A-NEXT: s_mov_b32 s10, 0xffffff
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f16_sdwa v2, v3, v0 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:DWORD
+; GFX90A-NEXT: v_and_or_b32 v2, v3, s10, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v1, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB6_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: s_addk_i32 s8, 0x200
+; GFX908-NEXT: s_and_b32 s8, s8, -4
+; GFX908-NEXT: v_mov_b32_e32 v5, s8
+; GFX908-NEXT: buffer_load_dword v2, v5, s[4:7], 0 offen
+; GFX908-NEXT: s_mov_b32 s10, 0xffffff
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f16_sdwa v1, v2, v0 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:DWORD
+; GFX908-NEXT: v_and_or_b32 v1, v2, s10, v1
+; GFX908-NEXT: v_mov_b32_e32 v4, v2
+; GFX908-NEXT: v_mov_b32_e32 v3, v1
+; GFX908-NEXT: buffer_atomic_cmpswap v[3:4], v5, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: v_mov_b32_e32 v2, v3
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB6_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 24, v3
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_addk_i32 s8, 0x200
+; GFX8-NEXT: s_and_b32 s8, s8, -4
+; GFX8-NEXT: v_mov_b32_e32 v5, s8
+; GFX8-NEXT: buffer_load_dword v2, v5, s[4:7], 0 offen
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f16_sdwa v1, v2, v0 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:DWORD
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffffff, v2
+; GFX8-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX8-NEXT: v_mov_b32_e32 v4, v2
+; GFX8-NEXT: v_mov_b32_e32 v3, v1
+; GFX8-NEXT: buffer_atomic_cmpswap v[3:4], v5, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v2, v3
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB6_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 24, v3
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_addk_i32 s8, 0x200
+; GFX7-NEXT: s_and_b32 s8, s8, -4
+; GFX7-NEXT: v_mov_b32_e32 v4, s8
+; GFX7-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v0
+; GFX7-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffffff, v1
+; GFX7-NEXT: v_add_f32_e32 v0, v0, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX7-NEXT: v_or_b32_e32 v0, v2, v0
+; GFX7-NEXT: v_mov_b32_e32 v3, v1
+; GFX7-NEXT: v_mov_b32_e32 v2, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB6_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_addk_i32 s8, 0x200
+; GFX6-NEXT: s_and_b32 s8, s8, -4
+; GFX6-NEXT: v_mov_b32_e32 v4, s8
+; GFX6-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v0
+; GFX6-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffffff, v1
+; GFX6-NEXT: v_add_f32_e32 v0, v0, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX6-NEXT: v_or_b32_e32 v0, v2, v0
+; GFX6-NEXT: v_mov_b32_e32 v3, v1
+; GFX6-NEXT: v_mov_b32_e32 v2, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB6_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fadd ptr addrspace(7) %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+define void @buffer_fat_ptr_agent_atomic_fadd_noret_f16__offset(ptr addrspace(7) inreg %ptr, half %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_addk_co_i32 s4, 0x200
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s4, s4, -4
+; GFX12-NEXT: v_mov_b32_e32 v3, s4
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: buffer_load_b32 v2, v3, s[0:3], null offen
+; GFX12-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v1, 24, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f16_e32 v1, v1, v0
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v1, 24, v1
+; GFX12-NEXT: v_and_or_b32 v1, 0xffffff, v2, v1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v5, v2 :: v_dual_mov_b32 v4, v1
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX12-NEXT: v_mov_b32_e32 v2, v4
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB7_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_addk_i32 s4, 0x200
+; GFX940-NEXT: s_and_b32 s4, s4, -4
+; GFX940-NEXT: v_mov_b32_e32 v1, s4
+; GFX940-NEXT: buffer_load_dword v3, v1, s[0:3], 0 offen
+; GFX940-NEXT: s_mov_b32 s6, 0xffffff
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_add_f16_sdwa v2, v3, v0 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:DWORD
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_and_or_b32 v2, v3, s6, v2
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[2:3]
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v1, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB7_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_addk_i32 s4, 0x200
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s4, s4, -4
+; GFX11-NEXT: v_mov_b32_e32 v3, s4
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: buffer_load_b32 v2, v3, s[0:3], 0 offen
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v1, 24, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f16_e32 v1, v1, v0
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v1, 24, v1
+; GFX11-NEXT: v_and_or_b32 v1, 0xffffff, v2, v1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v5, v2 :: v_dual_mov_b32 v4, v1
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX11-NEXT: v_mov_b32_e32 v2, v4
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB7_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_addk_i32 s8, 0x200
+; GFX10-NEXT: s_and_b32 s8, s8, -4
+; GFX10-NEXT: v_mov_b32_e32 v3, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v2, v3, s[4:7], 0 offen
+; GFX10-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f16_sdwa v1, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:DWORD
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 24, v1
+; GFX10-NEXT: v_and_or_b32 v1, 0xffffff, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v5, v2
+; GFX10-NEXT: v_mov_b32_e32 v4, v1
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX10-NEXT: v_mov_b32_e32 v2, v4
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB7_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: s_addk_i32 s8, 0x200
+; GFX90A-NEXT: s_and_b32 s8, s8, -4
+; GFX90A-NEXT: v_mov_b32_e32 v1, s8
+; GFX90A-NEXT: buffer_load_dword v3, v1, s[4:7], 0 offen
+; GFX90A-NEXT: s_mov_b32 s10, 0xffffff
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f16_sdwa v2, v3, v0 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:DWORD
+; GFX90A-NEXT: v_and_or_b32 v2, v3, s10, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v1, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB7_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: s_addk_i32 s8, 0x200
+; GFX908-NEXT: s_and_b32 s8, s8, -4
+; GFX908-NEXT: v_mov_b32_e32 v3, s8
+; GFX908-NEXT: buffer_load_dword v2, v3, s[4:7], 0 offen
+; GFX908-NEXT: s_mov_b32 s10, 0xffffff
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f16_sdwa v1, v2, v0 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:DWORD
+; GFX908-NEXT: v_and_or_b32 v1, v2, s10, v1
+; GFX908-NEXT: v_mov_b32_e32 v5, v2
+; GFX908-NEXT: v_mov_b32_e32 v4, v1
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v2
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: v_mov_b32_e32 v2, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB7_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_addk_i32 s8, 0x200
+; GFX8-NEXT: s_and_b32 s8, s8, -4
+; GFX8-NEXT: v_mov_b32_e32 v3, s8
+; GFX8-NEXT: buffer_load_dword v2, v3, s[4:7], 0 offen
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f16_sdwa v1, v2, v0 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:DWORD
+; GFX8-NEXT: v_and_b32_e32 v4, 0xffffff, v2
+; GFX8-NEXT: v_or_b32_e32 v1, v4, v1
+; GFX8-NEXT: v_mov_b32_e32 v5, v2
+; GFX8-NEXT: v_mov_b32_e32 v4, v1
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v2
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v2, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB7_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_addk_i32 s8, 0x200
+; GFX7-NEXT: s_and_b32 s8, s8, -4
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v0
+; GFX7-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffffff, v1
+; GFX7-NEXT: v_add_f32_e32 v0, v0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX7-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX7-NEXT: v_mov_b32_e32 v5, v1
+; GFX7-NEXT: v_mov_b32_e32 v4, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v1, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB7_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_addk_i32 s8, 0x200
+; GFX6-NEXT: s_and_b32 s8, s8, -4
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v0
+; GFX6-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffffff, v1
+; GFX6-NEXT: v_add_f32_e32 v0, v0, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX6-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX6-NEXT: v_mov_b32_e32 v5, v1
+; GFX6-NEXT: v_mov_b32_e32 v4, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v1, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB7_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(7) %ptr, i32 256
+ %unused = atomicrmw fadd ptr addrspace(7) %gep, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define half @buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset__waterfall(ptr addrspace(7) %ptr, half %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset__waterfall:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v4, 0x200, v4
+; GFX12-NEXT: s_mov_b32 s1, exec_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v4, -4, v4
+; GFX12-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: buffer_load_b32 v7, v4, s[4:7], null offen
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB8_1
+; GFX12-NEXT: ; %bb.2:
+; GFX12-NEXT: s_mov_b32 exec_lo, s1
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Loop Header: Depth=1
+; GFX12-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v6, 24, v7
+; GFX12-NEXT: s_mov_b32 s2, exec_lo
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f16_e32 v6, v6, v5
+; GFX12-NEXT: v_and_b32_e32 v6, 0xffff, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v6, 24, v6
+; GFX12-NEXT: v_and_or_b32 v6, 0xffffff, v7, v6
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v9, v7 :: v_dual_mov_b32 v8, v6
+; GFX12-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX12-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[8:9], v4, s[4:7], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB8_4
+; GFX12-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX12-NEXT: s_mov_b32 exec_lo, s2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v8, v7
+; GFX12-NEXT: v_mov_b32_e32 v7, v8
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB8_3
+; GFX12-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v8
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset__waterfall:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v4, 0x200, v4
+; GFX940-NEXT: v_and_b32_e32 v4, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v10, 0xffffff
+; GFX940-NEXT: s_mov_b64 s[2:3], exec
+; GFX940-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: buffer_load_dword v7, v4, s[4:7], 0 offen
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB8_1
+; GFX940-NEXT: ; %bb.2:
+; GFX940-NEXT: s_mov_b64 exec, s[2:3]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Loop Header: Depth=1
+; GFX940-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_add_f16_sdwa v6, v7, v5 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:DWORD
+; GFX940-NEXT: s_mov_b64 s[8:9], exec
+; GFX940-NEXT: v_and_or_b32 v6, v7, v10, v6
+; GFX940-NEXT: v_mov_b64_e32 v[8:9], v[6:7]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX940-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_atomic_cmpswap v[8:9], v4, s[4:7], 0 offen sc0
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB8_4
+; GFX940-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX940-NEXT: s_mov_b64 exec, s[8:9]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v7, v8
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB8_3
+; GFX940-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 24, v8
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset__waterfall:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v4, 0x200, v4
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v4, -4, v4
+; GFX11-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: buffer_load_b32 v7, v4, s[4:7], 0 offen
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB8_1
+; GFX11-NEXT: ; %bb.2:
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Loop Header: Depth=1
+; GFX11-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v6, 24, v7
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f16_e32 v6, v6, v5
+; GFX11-NEXT: v_and_b32_e32 v6, 0xffff, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v6, 24, v6
+; GFX11-NEXT: v_and_or_b32 v6, 0xffffff, v7, v6
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v9, v7 :: v_dual_mov_b32 v8, v6
+; GFX11-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX11-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[8:9], v4, s[4:7], 0 offen glc
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB8_4
+; GFX11-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v8, v7
+; GFX11-NEXT: v_mov_b32_e32 v7, v8
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB8_3
+; GFX11-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v8
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset__waterfall:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v4, 0x200, v4
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: v_and_b32_e32 v4, -4, v4
+; GFX10-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB8_1
+; GFX10-NEXT: ; %bb.2:
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Loop Header: Depth=1
+; GFX10-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_add_f16_sdwa v6, v7, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:DWORD
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_and_b32_e32 v6, 0xffff, v6
+; GFX10-NEXT: v_lshlrev_b32_e32 v6, 24, v6
+; GFX10-NEXT: v_and_or_b32 v6, 0xffffff, v7, v6
+; GFX10-NEXT: v_mov_b32_e32 v9, v7
+; GFX10-NEXT: v_mov_b32_e32 v8, v6
+; GFX10-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX10-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_atomic_cmpswap v[8:9], v4, s[8:11], 0 offen glc
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB8_4
+; GFX10-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v8, v7
+; GFX10-NEXT: v_mov_b32_e32 v7, v8
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB8_3
+; GFX10-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 24, v8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset__waterfall:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v4, 0x200, v4
+; GFX90A-NEXT: v_and_b32_e32 v4, -4, v4
+; GFX90A-NEXT: v_mov_b32_e32 v10, 0xffffff
+; GFX90A-NEXT: s_mov_b64 s[6:7], exec
+; GFX90A-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_nop 0
+; GFX90A-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB8_1
+; GFX90A-NEXT: ; %bb.2:
+; GFX90A-NEXT: s_mov_b64 exec, s[6:7]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Loop Header: Depth=1
+; GFX90A-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_add_f16_sdwa v6, v7, v5 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:DWORD
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v10, v6
+; GFX90A-NEXT: s_mov_b64 s[12:13], exec
+; GFX90A-NEXT: v_pk_mov_b32 v[8:9], v[6:7], v[6:7] op_sel:[0,1]
+; GFX90A-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX90A-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_atomic_cmpswap v[8:9], v4, s[8:11], 0 offen glc
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB8_4
+; GFX90A-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX90A-NEXT: s_mov_b64 exec, s[12:13]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v7, v8
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB8_3
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 24, v8
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset__waterfall:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v4, 0x200, v4
+; GFX908-NEXT: v_and_b32_e32 v4, -4, v4
+; GFX908-NEXT: v_mov_b32_e32 v10, 0xffffff
+; GFX908-NEXT: s_mov_b64 s[6:7], exec
+; GFX908-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_nop 0
+; GFX908-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_1
+; GFX908-NEXT: ; %bb.2:
+; GFX908-NEXT: s_mov_b64 exec, s[6:7]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Loop Header: Depth=1
+; GFX908-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_add_f16_sdwa v6, v7, v5 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:DWORD
+; GFX908-NEXT: v_and_or_b32 v6, v7, v10, v6
+; GFX908-NEXT: v_mov_b32_e32 v9, v7
+; GFX908-NEXT: s_mov_b64 s[12:13], exec
+; GFX908-NEXT: v_mov_b32_e32 v8, v6
+; GFX908-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX908-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_atomic_cmpswap v[8:9], v4, s[8:11], 0 offen glc
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_4
+; GFX908-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX908-NEXT: s_mov_b64 exec, s[12:13]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v7, v8
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_3
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 24, v8
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset__waterfall:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x200, v4
+; GFX8-NEXT: v_and_b32_e32 v4, -4, v4
+; GFX8-NEXT: s_mov_b64 s[6:7], exec
+; GFX8-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_nop 0
+; GFX8-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_1
+; GFX8-NEXT: ; %bb.2:
+; GFX8-NEXT: s_mov_b64 exec, s[6:7]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Loop Header: Depth=1
+; GFX8-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f16_sdwa v6, v7, v5 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:DWORD
+; GFX8-NEXT: v_and_b32_e32 v8, 0xffffff, v7
+; GFX8-NEXT: v_or_b32_e32 v6, v8, v6
+; GFX8-NEXT: v_mov_b32_e32 v9, v7
+; GFX8-NEXT: s_mov_b64 s[12:13], exec
+; GFX8-NEXT: v_mov_b32_e32 v8, v6
+; GFX8-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX8-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_atomic_cmpswap v[8:9], v4, s[8:11], 0 offen glc
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_4
+; GFX8-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX8-NEXT: s_mov_b64 exec, s[12:13]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v8, v7
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v7, v8
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_3
+; GFX8-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 24, v8
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset__waterfall:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x200, v4
+; GFX7-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB8_1
+; GFX7-NEXT: ; %bb.2:
+; GFX7-NEXT: s_mov_b64 exec, s[6:7]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v5
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v8, v4
+; GFX7-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Loop Header: Depth=1
+; GFX7-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffffff, v6
+; GFX7-NEXT: s_mov_b64 s[12:13], exec
+; GFX7-NEXT: v_add_f32_e32 v4, v4, v8
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX7-NEXT: v_or_b32_e32 v5, v5, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: v_mov_b32_e32 v5, v6
+; GFX7-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX7-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB8_4
+; GFX7-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX7-NEXT: s_mov_b64 exec, s[12:13]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB8_3
+; GFX7-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset__waterfall:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x200, v4
+; GFX6-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB8_1
+; GFX6-NEXT: ; %bb.2:
+; GFX6-NEXT: s_mov_b64 exec, s[6:7]
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v5
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v8, v4
+; GFX6-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Loop Header: Depth=1
+; GFX6-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffffff, v6
+; GFX6-NEXT: s_mov_b64 s[12:13], exec
+; GFX6-NEXT: v_add_f32_e32 v4, v4, v8
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX6-NEXT: v_or_b32_e32 v5, v5, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: v_mov_b32_e32 v5, v6
+; GFX6-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX6-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB8_4
+; GFX6-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX6-NEXT: s_mov_b64 exec, s[12:13]
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB8_3
+; GFX6-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fadd ptr addrspace(7) %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+; --------------------------------------------------------------------
+; bfloat
+; --------------------------------------------------------------------
+
+define bfloat @buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset(ptr addrspace(7) inreg %ptr, bfloat %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_addk_co_i32 s4, 0x200
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX12-NEXT: s_and_b32 s4, s4, -4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_mov_b32_e32 v4, s4
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: buffer_load_b32 v1, v4, s[0:3], null offen
+; GFX12-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX12-NEXT: v_add_f32_e32 v0, v0, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v3, 0x400000, v0
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX12-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX12-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[2:3], v4, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB9_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_addk_i32 s4, 0x200
+; GFX940-NEXT: s_and_b32 s4, s4, -4
+; GFX940-NEXT: v_mov_b32_e32 v4, s4
+; GFX940-NEXT: buffer_load_dword v1, v4, s[0:3], 0 offen
+; GFX940-NEXT: s_mov_b32 s6, 0xffffff
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX940-NEXT: s_movk_i32 s7, 0x7fff
+; GFX940-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX940-NEXT: v_add_f32_e32 v0, v0, v5
+; GFX940-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v3, 0x400000, v0
+; GFX940-NEXT: v_add3_u32 v2, v2, v0, s7
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX940-NEXT: v_and_or_b32 v0, v1, s6, v0
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], v[0:1]
+; GFX940-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB9_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_addk_i32 s4, 0x200
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX11-NEXT: s_and_b32 s4, s4, -4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v4, s4
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: buffer_load_b32 v1, v4, s[0:3], 0 offen
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT: v_add_f32_e32 v0, v0, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v3, 0x400000, v0
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX11-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[2:3], v4, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB9_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_addk_i32 s8, 0x200
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX10-NEXT: s_and_b32 s8, s8, -4
+; GFX10-NEXT: v_mov_b32_e32 v4, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX10-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX10-NEXT: v_add_f32_e32 v0, v0, v5
+; GFX10-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v3, 0x400000, v0
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX10-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX10-NEXT: v_mov_b32_e32 v3, v1
+; GFX10-NEXT: v_mov_b32_e32 v2, v0
+; GFX10-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB9_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: s_addk_i32 s8, 0x200
+; GFX90A-NEXT: s_and_b32 s8, s8, -4
+; GFX90A-NEXT: v_mov_b32_e32 v4, s8
+; GFX90A-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX90A-NEXT: s_mov_b32 s10, 0xffffff
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX90A-NEXT: s_movk_i32 s11, 0x7fff
+; GFX90A-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX90A-NEXT: v_add_f32_e32 v0, v0, v5
+; GFX90A-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v3, 0x400000, v0
+; GFX90A-NEXT: v_add3_u32 v2, v2, v0, s11
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX90A-NEXT: v_and_or_b32 v0, v1, s10, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB9_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: s_addk_i32 s8, 0x200
+; GFX908-NEXT: s_and_b32 s8, s8, -4
+; GFX908-NEXT: v_mov_b32_e32 v4, s8
+; GFX908-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX908-NEXT: s_mov_b32 s10, 0xffffff
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX908-NEXT: s_movk_i32 s11, 0x7fff
+; GFX908-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX908-NEXT: v_add_f32_e32 v0, v0, v5
+; GFX908-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v3, 0x400000, v0
+; GFX908-NEXT: v_add3_u32 v2, v2, v0, s11
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX908-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX908-NEXT: v_and_or_b32 v0, v1, s10, v0
+; GFX908-NEXT: v_mov_b32_e32 v3, v1
+; GFX908-NEXT: v_mov_b32_e32 v2, v0
+; GFX908-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB9_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_addk_i32 s8, 0x200
+; GFX8-NEXT: s_and_b32 s8, s8, -4
+; GFX8-NEXT: v_mov_b32_e32 v4, s8
+; GFX8-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX8-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: v_add_f32_e32 v0, v0, v5
+; GFX8-NEXT: v_bfe_u32 v3, v0, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v0
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
+; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v0, v3, v6, vcc
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffffff, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX8-NEXT: v_or_b32_e32 v0, v2, v0
+; GFX8-NEXT: v_mov_b32_e32 v3, v1
+; GFX8-NEXT: v_mov_b32_e32 v2, v0
+; GFX8-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB9_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_addk_i32 s8, 0x200
+; GFX7-NEXT: s_and_b32 s8, s8, -4
+; GFX7-NEXT: v_mov_b32_e32 v4, s8
+; GFX7-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v0
+; GFX7-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: v_add_f32_e32 v0, v0, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffffff, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX7-NEXT: v_or_b32_e32 v0, v2, v0
+; GFX7-NEXT: v_mov_b32_e32 v3, v1
+; GFX7-NEXT: v_mov_b32_e32 v2, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB9_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_addk_i32 s8, 0x200
+; GFX6-NEXT: s_and_b32 s8, s8, -4
+; GFX6-NEXT: v_mov_b32_e32 v4, s8
+; GFX6-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v0
+; GFX6-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: v_add_f32_e32 v0, v0, v5
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffffff, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX6-NEXT: v_or_b32_e32 v0, v2, v0
+; GFX6-NEXT: v_mov_b32_e32 v3, v1
+; GFX6-NEXT: v_mov_b32_e32 v2, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB9_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fadd ptr addrspace(7) %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+define void @buffer_fat_ptr_agent_atomic_fadd_noret_bf16__offset(ptr addrspace(7) inreg %ptr, bfloat %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_addk_co_i32 s4, 0x200
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX12-NEXT: s_and_b32 s4, s4, -4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_mov_b32_e32 v2, s4
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: buffer_load_b32 v1, v2, s[0:3], null offen
+; GFX12-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX12-NEXT: v_add_f32_e32 v0, v0, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v4, v0, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX12-NEXT: v_add3_u32 v4, v4, v0, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX12-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v2, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v4
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB10_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_addk_i32 s4, 0x200
+; GFX940-NEXT: s_and_b32 s4, s4, -4
+; GFX940-NEXT: v_mov_b32_e32 v2, s4
+; GFX940-NEXT: buffer_load_dword v1, v2, s[0:3], 0 offen
+; GFX940-NEXT: s_mov_b32 s6, 0xffffff
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX940-NEXT: s_movk_i32 s7, 0x7fff
+; GFX940-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX940-NEXT: v_add_f32_e32 v0, v0, v3
+; GFX940-NEXT: v_bfe_u32 v4, v0, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX940-NEXT: v_add3_u32 v4, v4, v0, s7
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX940-NEXT: v_and_or_b32 v0, v1, s6, v0
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[0:1]
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB10_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_addk_i32 s4, 0x200
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX11-NEXT: s_and_b32 s4, s4, -4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v2, s4
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: buffer_load_b32 v1, v2, s[0:3], 0 offen
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT: v_add_f32_e32 v0, v0, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v4, v0, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT: v_add3_u32 v4, v4, v0, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX11-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v2, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v4
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB10_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_addk_i32 s8, 0x200
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX10-NEXT: s_and_b32 s8, s8, -4
+; GFX10-NEXT: v_mov_b32_e32 v2, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX10-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX10-NEXT: v_add_f32_e32 v0, v0, v3
+; GFX10-NEXT: v_bfe_u32 v4, v0, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX10-NEXT: v_add3_u32 v4, v4, v0, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX10-NEXT: v_mov_b32_e32 v5, v1
+; GFX10-NEXT: v_mov_b32_e32 v4, v0
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v4
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB10_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: s_addk_i32 s8, 0x200
+; GFX90A-NEXT: s_and_b32 s8, s8, -4
+; GFX90A-NEXT: v_mov_b32_e32 v2, s8
+; GFX90A-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX90A-NEXT: s_mov_b32 s10, 0xffffff
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX90A-NEXT: s_movk_i32 s11, 0x7fff
+; GFX90A-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX90A-NEXT: v_add_f32_e32 v0, v0, v3
+; GFX90A-NEXT: v_bfe_u32 v4, v0, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX90A-NEXT: v_add3_u32 v4, v4, v0, s11
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX90A-NEXT: v_and_or_b32 v0, v1, s10, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB10_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: s_addk_i32 s8, 0x200
+; GFX908-NEXT: s_and_b32 s8, s8, -4
+; GFX908-NEXT: v_mov_b32_e32 v2, s8
+; GFX908-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX908-NEXT: s_mov_b32 s10, 0xffffff
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX908-NEXT: s_movk_i32 s11, 0x7fff
+; GFX908-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX908-NEXT: v_add_f32_e32 v0, v0, v3
+; GFX908-NEXT: v_bfe_u32 v4, v0, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX908-NEXT: v_add3_u32 v4, v4, v0, s11
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX908-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX908-NEXT: v_and_or_b32 v0, v1, s10, v0
+; GFX908-NEXT: v_mov_b32_e32 v5, v1
+; GFX908-NEXT: v_mov_b32_e32 v4, v0
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: v_mov_b32_e32 v1, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB10_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_addk_i32 s8, 0x200
+; GFX8-NEXT: s_and_b32 s8, s8, -4
+; GFX8-NEXT: v_mov_b32_e32 v2, s8
+; GFX8-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX8-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: v_add_f32_e32 v0, v0, v3
+; GFX8-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v0
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
+; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v0, v5, v6, vcc
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: v_and_b32_e32 v4, 0xffffff, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX8-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX8-NEXT: v_mov_b32_e32 v5, v1
+; GFX8-NEXT: v_mov_b32_e32 v4, v0
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v1, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB10_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_addk_i32 s8, 0x200
+; GFX7-NEXT: s_and_b32 s8, s8, -4
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX7-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: v_add_f32_e32 v0, v0, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffffff, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX7-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX7-NEXT: v_mov_b32_e32 v5, v1
+; GFX7-NEXT: v_mov_b32_e32 v4, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v1, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB10_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_addk_i32 s8, 0x200
+; GFX6-NEXT: s_and_b32 s8, s8, -4
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX6-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: v_add_f32_e32 v0, v0, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffffff, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX6-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX6-NEXT: v_mov_b32_e32 v5, v1
+; GFX6-NEXT: v_mov_b32_e32 v4, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v1, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB10_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(7) %ptr, i32 256
+ %unused = atomicrmw fadd ptr addrspace(7) %gep, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define bfloat @buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__waterfall(ptr addrspace(7) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__waterfall:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v4, 0x200, v4
+; GFX12-NEXT: s_mov_b32 s1, exec_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX12-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: buffer_load_b32 v6, v7, s[4:7], null offen
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB11_1
+; GFX12-NEXT: ; %bb.2:
+; GFX12-NEXT: s_mov_b32 exec_lo, s1
+; GFX12-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Loop Header: Depth=1
+; GFX12-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX12-NEXT: s_mov_b32 s2, exec_lo
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: v_add_f32_e32 v4, v4, v8
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v4
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX12-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v4, v5, v9, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX12-NEXT: v_and_or_b32 v5, 0xffffff, v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_mov_b32_e32 v4, v5
+; GFX12-NEXT: v_mov_b32_e32 v5, v6
+; GFX12-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX12-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB11_4
+; GFX12-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX12-NEXT: s_mov_b32 exec_lo, s2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX12-NEXT: v_mov_b32_e32 v6, v4
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB11_3
+; GFX12-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__waterfall:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v4, 0x200, v4
+; GFX940-NEXT: v_and_b32_e32 v8, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v9, 0xffffff
+; GFX940-NEXT: s_mov_b64 s[2:3], exec
+; GFX940-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: buffer_load_dword v7, v8, s[4:7], 0 offen
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB11_1
+; GFX940-NEXT: ; %bb.2:
+; GFX940-NEXT: s_mov_b64 exec, s[2:3]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v10, 16, v5
+; GFX940-NEXT: s_movk_i32 s10, 0x7fff
+; GFX940-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Loop Header: Depth=1
+; GFX940-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, 24, v7
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX940-NEXT: v_add_f32_e32 v4, v4, v10
+; GFX940-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX940-NEXT: v_add3_u32 v5, v5, v4, s10
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_mov_b64 s[8:9], exec
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX940-NEXT: v_and_or_b32 v6, v7, v9, v4
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[6:7]
+; GFX940-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX940-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[4:7], 0 offen sc0
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB11_4
+; GFX940-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX940-NEXT: s_mov_b64 exec, s[8:9]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB11_3
+; GFX940-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__waterfall:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v4, 0x200, v4
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX11-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: buffer_load_b32 v6, v7, s[4:7], 0 offen
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB11_1
+; GFX11-NEXT: ; %bb.2:
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Loop Header: Depth=1
+; GFX11-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: v_add_f32_e32 v4, v4, v8
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v4
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v4, v5, v9, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX11-NEXT: v_and_or_b32 v5, 0xffffff, v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_mov_b32_e32 v4, v5
+; GFX11-NEXT: v_mov_b32_e32 v5, v6
+; GFX11-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX11-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], 0 offen glc
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB11_4
+; GFX11-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX11-NEXT: v_mov_b32_e32 v6, v4
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB11_3
+; GFX11-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__waterfall:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v4, 0x200, v4
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX10-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB11_1
+; GFX10-NEXT: ; %bb.2:
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX10-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Loop Header: Depth=1
+; GFX10-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX10-NEXT: v_add_f32_e32 v4, v4, v8
+; GFX10-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v4
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX10-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v4, v5, v9, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX10-NEXT: v_and_or_b32 v5, 0xffffff, v6, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v5
+; GFX10-NEXT: v_mov_b32_e32 v5, v6
+; GFX10-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX10-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB11_4
+; GFX10-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB11_3
+; GFX10-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__waterfall:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v4, 0x200, v4
+; GFX90A-NEXT: v_and_b32_e32 v8, -4, v4
+; GFX90A-NEXT: v_mov_b32_e32 v9, 0xffffff
+; GFX90A-NEXT: s_mov_b64 s[6:7], exec
+; GFX90A-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_nop 0
+; GFX90A-NEXT: buffer_load_dword v7, v8, s[8:11], 0 offen
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_1
+; GFX90A-NEXT: ; %bb.2:
+; GFX90A-NEXT: s_mov_b64 exec, s[6:7]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v10, 16, v5
+; GFX90A-NEXT: s_movk_i32 s14, 0x7fff
+; GFX90A-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Loop Header: Depth=1
+; GFX90A-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, 24, v7
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX90A-NEXT: v_add_f32_e32 v4, v4, v10
+; GFX90A-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX90A-NEXT: v_add3_u32 v5, v5, v4, s14
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v9, v4
+; GFX90A-NEXT: s_mov_b64 s[12:13], exec
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[6:7] op_sel:[0,1]
+; GFX90A-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX90A-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[8:11], 0 offen glc
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_4
+; GFX90A-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX90A-NEXT: s_mov_b64 exec, s[12:13]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_3
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__waterfall:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v4, 0x200, v4
+; GFX908-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX908-NEXT: v_mov_b32_e32 v8, 0xffffff
+; GFX908-NEXT: s_mov_b64 s[6:7], exec
+; GFX908-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_nop 0
+; GFX908-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_1
+; GFX908-NEXT: ; %bb.2:
+; GFX908-NEXT: s_mov_b64 exec, s[6:7]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v9, 16, v5
+; GFX908-NEXT: s_movk_i32 s14, 0x7fff
+; GFX908-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Loop Header: Depth=1
+; GFX908-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX908-NEXT: v_add_f32_e32 v4, v4, v9
+; GFX908-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX908-NEXT: v_add3_u32 v5, v5, v4, s14
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v4
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX908-NEXT: v_cndmask_b32_e32 v4, v5, v10, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX908-NEXT: v_and_or_b32 v5, v6, v8, v4
+; GFX908-NEXT: v_mov_b32_e32 v4, v5
+; GFX908-NEXT: s_mov_b64 s[12:13], exec
+; GFX908-NEXT: v_mov_b32_e32 v5, v6
+; GFX908-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX908-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_4
+; GFX908-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX908-NEXT: s_mov_b64 exec, s[12:13]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_3
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__waterfall:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x200, v4
+; GFX8-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX8-NEXT: s_mov_b64 s[6:7], exec
+; GFX8-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_nop 0
+; GFX8-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_1
+; GFX8-NEXT: ; %bb.2:
+; GFX8-NEXT: s_mov_b64 exec, s[6:7]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX8-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Loop Header: Depth=1
+; GFX8-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX8-NEXT: v_add_f32_e32 v4, v4, v8
+; GFX8-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v4
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v4, v5, v9, vcc
+; GFX8-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffffff, v6
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v4
+; GFX8-NEXT: v_mov_b32_e32 v4, v5
+; GFX8-NEXT: s_mov_b64 s[12:13], exec
+; GFX8-NEXT: v_mov_b32_e32 v5, v6
+; GFX8-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX8-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_4
+; GFX8-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX8-NEXT: s_mov_b64 exec, s[12:13]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_3
+; GFX8-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__waterfall:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x200, v4
+; GFX7-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB11_1
+; GFX7-NEXT: ; %bb.2:
+; GFX7-NEXT: s_mov_b64 exec, s[6:7]
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v5
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v4
+; GFX7-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Loop Header: Depth=1
+; GFX7-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_add_f32_e32 v4, v4, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffffff, v6
+; GFX7-NEXT: v_or_b32_e32 v5, v5, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_mov_b64 s[12:13], exec
+; GFX7-NEXT: v_mov_b32_e32 v5, v6
+; GFX7-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX7-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB11_4
+; GFX7-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX7-NEXT: s_mov_b64 exec, s[12:13]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB11_3
+; GFX7-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__waterfall:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x200, v4
+; GFX6-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB11_1
+; GFX6-NEXT: ; %bb.2:
+; GFX6-NEXT: s_mov_b64 exec, s[6:7]
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v5
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: v_and_b32_e32 v8, 0xffff0000, v4
+; GFX6-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Loop Header: Depth=1
+; GFX6-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_add_f32_e32 v4, v4, v8
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffffff, v6
+; GFX6-NEXT: v_or_b32_e32 v5, v5, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_mov_b64 s[12:13], exec
+; GFX6-NEXT: v_mov_b32_e32 v5, v6
+; GFX6-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX6-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB11_4
+; GFX6-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX6-NEXT: s_mov_b64 exec, s[12:13]
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB11_3
+; GFX6-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fadd ptr addrspace(7) %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+; --------------------------------------------------------------------
+; <2 x half>
+; --------------------------------------------------------------------
+
+define <2 x half> @buffer_fat_ptr_agent_atomic_fadd_ret_v2f16__offset(ptr addrspace(7) inreg %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v0
+; GFX12-NEXT: v_mov_b32_e32 v0, s4
+; GFX12-NEXT: s_addk_co_i32 s4, 0x400
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_mov_b32_e32 v3, s4
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: buffer_load_b32 v0, v0, s[0:3], null offen offset:1024
+; GFX12-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v5, v0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v3, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB12_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v0
+; GFX940-NEXT: v_mov_b32_e32 v0, s4
+; GFX940-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen offset:1024
+; GFX940-NEXT: s_add_i32 s6, s4, 0x400
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_mov_b32_e32 v3, s6
+; GFX940-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v0
+; GFX940-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_mov_b64_e32 v[0:1], v[4:5]
+; GFX940-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB12_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v0
+; GFX11-NEXT: v_mov_b32_e32 v0, s4
+; GFX11-NEXT: s_addk_i32 s4, 0x400
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v3, s4
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
+; GFX11-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v5, v0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v3, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB12_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v0
+; GFX10-NEXT: v_mov_b32_e32 v0, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: v_mov_b32_e32 v3, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v5, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB12_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v0
+; GFX90A-NEXT: v_mov_b32_e32 v0, s8
+; GFX90A-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX90A-NEXT: s_add_i32 s10, s8, 0x400
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_mov_b32_e32 v3, s10
+; GFX90A-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v0
+; GFX90A-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB12_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v0
+; GFX908-NEXT: v_mov_b32_e32 v0, s8
+; GFX908-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_add_i32 s10, s8, 0x400
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_mov_b32_e32 v3, s10
+; GFX908-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v5, v0
+; GFX908-NEXT: v_pk_add_f16 v4, v5, v2
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB12_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v0
+; GFX8-NEXT: v_mov_b32_e32 v0, s8
+; GFX8-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_add_i32 s10, s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_mov_b32_e32 v3, s10
+; GFX8-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v5, v0
+; GFX8-NEXT: v_add_f16_sdwa v0, v5, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v1, v5, v2
+; GFX8-NEXT: v_or_b32_e32 v4, v1, v0
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: v_mov_b32_e32 v1, v5
+; GFX8-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB12_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dword v3, v2, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v0
+; GFX7-NEXT: s_add_i32 s10, s8, 0x400
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, s10
+; GFX7-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v0, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v0
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v4, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB12_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dword v3, v2, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v0
+; GFX6-NEXT: s_add_i32 s10, s8, 0x400
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, s10
+; GFX6-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v6
+; GFX6-NEXT: v_or_b32_e32 v6, v0, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v5
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v0
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v4, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB12_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fadd ptr addrspace(7) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define void @buffer_fat_ptr_agent_atomic_fadd_noret_v2f16__offset(ptr addrspace(7) inreg %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_v2f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v1, s4
+; GFX12-NEXT: s_addk_co_i32 s4, 0x400
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_mov_b32_e32 v3, s4
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: buffer_load_b32 v2, v1, s[0:3], null offen offset:1024
+; GFX12-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v1, v2, v0
+; GFX12-NEXT: v_mov_b32_e32 v5, v2
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_mov_b32_e32 v4, v1
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX12-NEXT: v_mov_b32_e32 v2, v4
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB13_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_v2f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, s4
+; GFX940-NEXT: buffer_load_dword v3, v1, s[0:3], 0 offen offset:1024
+; GFX940-NEXT: s_add_i32 s6, s4, 0x400
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_mov_b32_e32 v1, s6
+; GFX940-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v2, v3, v0
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[2:3]
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v1, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB13_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_v2f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v1, s4
+; GFX11-NEXT: s_addk_i32 s4, 0x400
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v3, s4
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: buffer_load_b32 v2, v1, s[0:3], 0 offen offset:1024
+; GFX11-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v1, v2, v0
+; GFX11-NEXT: v_mov_b32_e32 v5, v2
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_mov_b32_e32 v4, v1
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX11-NEXT: v_mov_b32_e32 v2, v4
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB13_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_v2f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: v_mov_b32_e32 v3, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v2, v1, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v1, v2, v0
+; GFX10-NEXT: v_mov_b32_e32 v5, v2
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_mov_b32_e32 v4, v1
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2
+; GFX10-NEXT: v_mov_b32_e32 v2, v4
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB13_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_v2f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, s8
+; GFX90A-NEXT: buffer_load_dword v3, v1, s[4:7], 0 offen offset:1024
+; GFX90A-NEXT: s_add_i32 s10, s8, 0x400
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_mov_b32_e32 v1, s10
+; GFX90A-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v2, v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v1, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB13_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_v2f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, s8
+; GFX908-NEXT: buffer_load_dword v2, v1, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_add_i32 s10, s8, 0x400
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_mov_b32_e32 v3, s10
+; GFX908-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v1, v2, v0
+; GFX908-NEXT: v_mov_b32_e32 v5, v2
+; GFX908-NEXT: v_mov_b32_e32 v4, v1
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v2
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: v_mov_b32_e32 v2, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB13_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_v2f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, s8
+; GFX8-NEXT: buffer_load_dword v2, v1, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_add_i32 s10, s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_mov_b32_e32 v3, s10
+; GFX8-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f16_sdwa v1, v2, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v4, v2, v0
+; GFX8-NEXT: v_or_b32_e32 v1, v4, v1
+; GFX8-NEXT: v_mov_b32_e32 v5, v2
+; GFX8-NEXT: v_mov_b32_e32 v4, v1
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v2
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v2, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB13_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_v2f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dword v2, v2, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v0
+; GFX7-NEXT: s_add_i32 s10, s8, 0x400
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v5
+; GFX7-NEXT: v_mov_b32_e32 v2, s10
+; GFX7-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_add_f32_e32 v5, v5, v0
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_or_b32_e32 v5, v3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX7-NEXT: v_or_b32_e32 v4, v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v2, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB13_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_v2f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dword v2, v2, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v0
+; GFX6-NEXT: s_add_i32 s10, s8, 0x400
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v5
+; GFX6-NEXT: v_mov_b32_e32 v2, s10
+; GFX6-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_add_f32_e32 v5, v5, v0
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_or_b32_e32 v5, v3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX6-NEXT: v_or_b32_e32 v4, v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v2, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB13_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(7) %ptr, i32 256
+ %unused = atomicrmw fadd ptr addrspace(7) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x half> @buffer_fat_ptr_agent_atomic_fadd_ret_v2f16__offset__waterfall(ptr addrspace(7) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2f16__offset__waterfall:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v9, 0x400, v4
+; GFX12-NEXT: s_mov_b32 s1, exec_lo
+; GFX12-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: buffer_load_b32 v8, v4, s[4:7], null offen offset:1024
+; GFX12-NEXT: ; implicit-def: $vgpr4
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB14_1
+; GFX12-NEXT: ; %bb.2:
+; GFX12-NEXT: s_mov_b32 exec_lo, s1
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Loop Header: Depth=1
+; GFX12-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_add_f16 v7, v8, v5
+; GFX12-NEXT: s_mov_b32 s2, exec_lo
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_mov_b32_e32 v6, v7
+; GFX12-NEXT: v_mov_b32_e32 v7, v8
+; GFX12-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX12-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[6:7], v9, s[4:7], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB14_4
+; GFX12-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX12-NEXT: s_mov_b32 exec_lo, s2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v6, v8
+; GFX12-NEXT: v_mov_b32_e32 v8, v6
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB14_3
+; GFX12-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v6
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2f16__offset__waterfall:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v10, 0x400, v4
+; GFX940-NEXT: s_mov_b64 s[2:3], exec
+; GFX940-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: buffer_load_dword v9, v4, s[4:7], 0 offen offset:1024
+; GFX940-NEXT: ; implicit-def: $vgpr4
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB14_1
+; GFX940-NEXT: ; %bb.2:
+; GFX940-NEXT: s_mov_b64 exec, s[2:3]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Loop Header: Depth=1
+; GFX940-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_add_f16 v8, v9, v5
+; GFX940-NEXT: s_mov_b64 s[8:9], exec
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[8:9]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX940-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_atomic_cmpswap v[6:7], v10, s[4:7], 0 offen sc0
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB14_4
+; GFX940-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX940-NEXT: s_mov_b64 exec, s[8:9]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v6, v9
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v9, v6
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB14_3
+; GFX940-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v6
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2f16__offset__waterfall:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v9, 0x400, v4
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: buffer_load_b32 v8, v4, s[4:7], 0 offen offset:1024
+; GFX11-NEXT: ; implicit-def: $vgpr4
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB14_1
+; GFX11-NEXT: ; %bb.2:
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Loop Header: Depth=1
+; GFX11-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_add_f16 v7, v8, v5
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_mov_b32_e32 v6, v7
+; GFX11-NEXT: v_mov_b32_e32 v7, v8
+; GFX11-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX11-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[6:7], v9, s[4:7], 0 offen glc
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB14_4
+; GFX11-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v6, v8
+; GFX11-NEXT: v_mov_b32_e32 v8, v6
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB14_3
+; GFX11-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v6
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2f16__offset__waterfall:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v9, 0x400, v4
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: buffer_load_dword v8, v4, s[8:11], 0 offen offset:1024
+; GFX10-NEXT: ; implicit-def: $vgpr4
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB14_1
+; GFX10-NEXT: ; %bb.2:
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Loop Header: Depth=1
+; GFX10-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_add_f16 v7, v8, v5
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_mov_b32_e32 v6, v7
+; GFX10-NEXT: v_mov_b32_e32 v7, v8
+; GFX10-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX10-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_atomic_cmpswap v[6:7], v9, s[8:11], 0 offen glc
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB14_4
+; GFX10-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v6, v8
+; GFX10-NEXT: v_mov_b32_e32 v8, v6
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB14_3
+; GFX10-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v6
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2f16__offset__waterfall:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v10, 0x400, v4
+; GFX90A-NEXT: s_mov_b64 s[6:7], exec
+; GFX90A-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_nop 0
+; GFX90A-NEXT: buffer_load_dword v9, v4, s[8:11], 0 offen offset:1024
+; GFX90A-NEXT: ; implicit-def: $vgpr4
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_1
+; GFX90A-NEXT: ; %bb.2:
+; GFX90A-NEXT: s_mov_b64 exec, s[6:7]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Loop Header: Depth=1
+; GFX90A-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_add_f16 v8, v9, v5
+; GFX90A-NEXT: s_mov_b64 s[12:13], exec
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[8:9], v[8:9] op_sel:[0,1]
+; GFX90A-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX90A-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_atomic_cmpswap v[6:7], v10, s[8:11], 0 offen glc
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_4
+; GFX90A-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX90A-NEXT: s_mov_b64 exec, s[12:13]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v6, v9
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v9, v6
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_3
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v6
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2f16__offset__waterfall:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v9, 0x400, v4
+; GFX908-NEXT: s_mov_b64 s[6:7], exec
+; GFX908-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_nop 0
+; GFX908-NEXT: buffer_load_dword v8, v4, s[8:11], 0 offen offset:1024
+; GFX908-NEXT: ; implicit-def: $vgpr4
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_1
+; GFX908-NEXT: ; %bb.2:
+; GFX908-NEXT: s_mov_b64 exec, s[6:7]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Loop Header: Depth=1
+; GFX908-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_add_f16 v7, v8, v5
+; GFX908-NEXT: v_mov_b32_e32 v6, v7
+; GFX908-NEXT: s_mov_b64 s[12:13], exec
+; GFX908-NEXT: v_mov_b32_e32 v7, v8
+; GFX908-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX908-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_atomic_cmpswap v[6:7], v9, s[8:11], 0 offen glc
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_4
+; GFX908-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX908-NEXT: s_mov_b64 exec, s[12:13]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v6, v8
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v8, v6
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_3
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v6
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2f16__offset__waterfall:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x400, v4
+; GFX8-NEXT: s_mov_b64 s[6:7], exec
+; GFX8-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_nop 0
+; GFX8-NEXT: buffer_load_dword v8, v4, s[8:11], 0 offen offset:1024
+; GFX8-NEXT: ; implicit-def: $vgpr4
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_1
+; GFX8-NEXT: ; %bb.2:
+; GFX8-NEXT: s_mov_b64 exec, s[6:7]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Loop Header: Depth=1
+; GFX8-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_add_f16_sdwa v4, v8, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_add_f16_e32 v6, v8, v5
+; GFX8-NEXT: v_or_b32_e32 v7, v6, v4
+; GFX8-NEXT: v_mov_b32_e32 v6, v7
+; GFX8-NEXT: s_mov_b64 s[12:13], exec
+; GFX8-NEXT: v_mov_b32_e32 v7, v8
+; GFX8-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX8-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_atomic_cmpswap v[6:7], v9, s[8:11], 0 offen glc
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_4
+; GFX8-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX8-NEXT: s_mov_b64 exec, s[12:13]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v6, v8
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v8, v6
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_3
+; GFX8-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v6
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2f16__offset__waterfall:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v9, vcc, 0x400, v4
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen offset:1024
+; GFX7-NEXT: ; implicit-def: $vgpr4
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_1
+; GFX7-NEXT: ; %bb.2:
+; GFX7-NEXT: s_mov_b64 exec, s[6:7]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v5
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v10, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v11, v8
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Loop Header: Depth=1
+; GFX7-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: s_mov_b64 s[12:13], exec
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v10
+; GFX7-NEXT: v_add_f32_e32 v7, v7, v11
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX7-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v9, s[8:11], 0 offen glc
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_4
+; GFX7-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX7-NEXT: s_mov_b64 exec, s[12:13]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_3
+; GFX7-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v0, v4
+; GFX7-NEXT: v_mov_b32_e32 v1, v5
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2f16__offset__waterfall:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v9, vcc, 0x400, v4
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen offset:1024
+; GFX6-NEXT: ; implicit-def: $vgpr4
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB14_1
+; GFX6-NEXT: ; %bb.2:
+; GFX6-NEXT: s_mov_b64 exec, s[6:7]
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v5
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v10, v6
+; GFX6-NEXT: v_cvt_f32_f16_e32 v11, v8
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Loop Header: Depth=1
+; GFX6-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: s_mov_b64 s[12:13], exec
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v10
+; GFX6-NEXT: v_add_f32_e32 v7, v7, v11
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX6-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX6-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v9, s[8:11], 0 offen glc
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB14_4
+; GFX6-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX6-NEXT: s_mov_b64 exec, s[12:13]
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB14_3
+; GFX6-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v0, v4
+; GFX6-NEXT: v_mov_b32_e32 v1, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fadd ptr addrspace(7) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+; --------------------------------------------------------------------
+; <2 x bfloat>
+; --------------------------------------------------------------------
+
+define <2 x bfloat> @buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset(ptr addrspace(7) inreg %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s4
+; GFX12-NEXT: s_addk_co_i32 s4, 0x400
+; GFX12-NEXT: s_mov_b32 s5, 0
+; GFX12-NEXT: v_mov_b32_e32 v4, s4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX12-NEXT: buffer_load_b32 v0, v0, s[0:3], null offen offset:1024
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX12-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX12-NEXT: v_add_f32_e32 v1, v1, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v1, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX12-NEXT: v_add3_u32 v7, v7, v1, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_cndmask_b32 v1, v7, v9 :: v_dual_lshlrev_b32 v0, 16, v6
+; GFX12-NEXT: v_add_f32_e32 v0, v0, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX12-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX12-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v0, v5, v8, s4
+; GFX12-NEXT: v_perm_b32 v5, v1, v0, 0x7060302
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v0, v5 :: v_dual_mov_b32 v1, v6
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v4, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX12-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5
+; GFX12-NEXT: s_cbranch_execnz .LBB15_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, v0
+; GFX940-NEXT: v_mov_b32_e32 v0, s4
+; GFX940-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen offset:1024
+; GFX940-NEXT: s_addk_i32 s4, 0x400
+; GFX940-NEXT: s_mov_b64 s[6:7], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX940-NEXT: s_movk_i32 s8, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX940-NEXT: s_mov_b32 s9, 0x7060302
+; GFX940-NEXT: v_mov_b32_e32 v4, s4
+; GFX940-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v7
+; GFX940-NEXT: v_add_f32_e32 v0, v0, v2
+; GFX940-NEXT: v_add_f32_e32 v1, v1, v3
+; GFX940-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v1, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v0
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX940-NEXT: v_add3_u32 v5, v5, v0, s8
+; GFX940-NEXT: v_add3_u32 v8, v8, v1, s8
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX940-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_cndmask_b32_e32 v1, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v0, v5, v6, s[4:5]
+; GFX940-NEXT: v_perm_b32 v6, v1, v0, s9
+; GFX940-NEXT: v_mov_b64_e32 v[0:1], v[6:7]
+; GFX940-NEXT: buffer_atomic_cmpswap v[0:1], v4, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v7
+; GFX940-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX940-NEXT: s_cbranch_execnz .LBB15_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s4
+; GFX11-NEXT: s_addk_i32 s4, 0x400
+; GFX11-NEXT: s_mov_b32 s5, 0
+; GFX11-NEXT: v_mov_b32_e32 v4, s4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX11-NEXT: buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX11-NEXT: v_add_f32_e32 v1, v1, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v1, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT: v_add3_u32 v7, v7, v1, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_cndmask_b32 v1, v7, v9 :: v_dual_lshlrev_b32 v0, 16, v6
+; GFX11-NEXT: v_add_f32_e32 v0, v0, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX11-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX11-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v0, v5, v8, s4
+; GFX11-NEXT: v_perm_b32 v5, v1, v0, 0x7060302
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v0, v5 :: v_dual_mov_b32 v1, v6
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v4, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX11-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5
+; GFX11-NEXT: s_cbranch_execnz .LBB15_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_mov_b32_e32 v0, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: s_mov_b32 s9, 0
+; GFX10-NEXT: v_mov_b32_e32 v4, s8
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX10-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX10-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX10-NEXT: v_add_f32_e32 v0, v0, v2
+; GFX10-NEXT: v_add_f32_e32 v1, v1, v3
+; GFX10-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v1, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX10-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v1, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s8, v0, v0
+; GFX10-NEXT: v_cndmask_b32_e32 v1, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v0, v5, v8, s8
+; GFX10-NEXT: v_perm_b32 v5, v1, v0, 0x7060302
+; GFX10-NEXT: v_mov_b32_e32 v0, v5
+; GFX10-NEXT: v_mov_b32_e32 v1, v6
+; GFX10-NEXT: buffer_atomic_cmpswap v[0:1], v4, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s9, vcc_lo, s9
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s9
+; GFX10-NEXT: s_cbranch_execnz .LBB15_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s9
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: v_mov_b32_e32 v0, s8
+; GFX90A-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX90A-NEXT: s_addk_i32 s8, 0x400
+; GFX90A-NEXT: s_mov_b64 s[10:11], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX90A-NEXT: s_movk_i32 s12, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX90A-NEXT: s_mov_b32 s13, 0x7060302
+; GFX90A-NEXT: v_mov_b32_e32 v4, s8
+; GFX90A-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v7
+; GFX90A-NEXT: v_add_f32_e32 v0, v0, v2
+; GFX90A-NEXT: v_add_f32_e32 v1, v1, v3
+; GFX90A-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v1, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v0
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX90A-NEXT: v_add3_u32 v5, v5, v0, s12
+; GFX90A-NEXT: v_add3_u32 v8, v8, v1, s12
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[8:9], v0, v0
+; GFX90A-NEXT: v_cndmask_b32_e64 v0, v5, v6, s[8:9]
+; GFX90A-NEXT: v_cndmask_b32_e32 v1, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v1, v0, s13
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[6:7] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[0:1], v4, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v7
+; GFX90A-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX90A-NEXT: s_cbranch_execnz .LBB15_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: v_mov_b32_e32 v0, s8
+; GFX908-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_addk_i32 s8, 0x400
+; GFX908-NEXT: s_mov_b64 s[10:11], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX908-NEXT: s_movk_i32 s12, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX908-NEXT: s_mov_b32 s13, 0x7060302
+; GFX908-NEXT: v_mov_b32_e32 v4, s8
+; GFX908-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX908-NEXT: v_add_f32_e32 v0, v0, v2
+; GFX908-NEXT: v_add_f32_e32 v1, v1, v3
+; GFX908-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v1, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX908-NEXT: v_add3_u32 v5, v5, v0, s12
+; GFX908-NEXT: v_add3_u32 v8, v8, v1, s12
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX908-NEXT: v_cmp_u_f32_e64 s[8:9], v0, v0
+; GFX908-NEXT: v_cndmask_b32_e64 v0, v5, v7, s[8:9]
+; GFX908-NEXT: v_cndmask_b32_e32 v1, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v5, v1, v0, s13
+; GFX908-NEXT: v_mov_b32_e32 v0, v5
+; GFX908-NEXT: v_mov_b32_e32 v1, v6
+; GFX908-NEXT: buffer_atomic_cmpswap v[0:1], v4, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX908-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX908-NEXT: s_cbranch_execnz .LBB15_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_mov_b32_e32 v0, s8
+; GFX8-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_addk_i32 s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[10:11], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX8-NEXT: v_mov_b32_e32 v4, s8
+; GFX8-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX8-NEXT: v_add_f32_e32 v0, v0, v2
+; GFX8-NEXT: v_add_f32_e32 v1, v1, v3
+; GFX8-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v1, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v0
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[8:9], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v1, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v5, v7, s[8:9]
+; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX8-NEXT: v_alignbit_b32 v5, v1, v0, 16
+; GFX8-NEXT: v_mov_b32_e32 v0, v5
+; GFX8-NEXT: v_mov_b32_e32 v1, v6
+; GFX8-NEXT: buffer_atomic_cmpswap v[0:1], v4, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX8-NEXT: s_cbranch_execnz .LBB15_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dword v4, v2, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: s_addk_i32 s8, 0x400
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: s_mov_b64 s[10:11], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v0
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, s8
+; GFX7-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v3
+; GFX7-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v0, v0, v5, 16
+; GFX7-NEXT: v_mov_b32_e32 v6, v1
+; GFX7-NEXT: v_mov_b32_e32 v5, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v4, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v5
+; GFX7-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX7-NEXT: s_cbranch_execnz .LBB15_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dword v4, v2, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: s_addk_i32 s8, 0x400
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: s_mov_b64 s[10:11], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v0
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, s8
+; GFX6-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v0
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v3
+; GFX6-NEXT: v_add_f32_e32 v5, v5, v2
+; GFX6-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v0, v0, v5, 16
+; GFX6-NEXT: v_mov_b32_e32 v6, v1
+; GFX6-NEXT: v_mov_b32_e32 v5, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v4, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v1
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v5
+; GFX6-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX6-NEXT: s_cbranch_execnz .LBB15_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fadd ptr addrspace(7) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @buffer_fat_ptr_agent_atomic_fadd_noret_v2bf16__offset(ptr addrspace(7) inreg %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_v2bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v1, s4 :: v_dual_lshlrev_b32 v2, 16, v0
+; GFX12-NEXT: s_addk_co_i32 s4, 0x400
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_and_b32 v3, 0xffff0000, v0
+; GFX12-NEXT: buffer_load_b32 v1, v1, s[0:3], null offen offset:1024
+; GFX12-NEXT: s_mov_b32 s5, 0
+; GFX12-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_add_f32 v5, v5, v3 :: v_dual_add_f32 v0, v0, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_add3_u32 v6, v6, v0, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v0, v6, v8, s4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v0, v5, v0, 0x7060302
+; GFX12-NEXT: v_dual_mov_b32 v6, v1 :: v_dual_mov_b32 v5, v0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[5:6], v4, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v5
+; GFX12-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5
+; GFX12-NEXT: s_cbranch_execnz .LBB16_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_v2bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, s4
+; GFX940-NEXT: buffer_load_dword v1, v1, s[0:3], 0 offen offset:1024
+; GFX940-NEXT: s_addk_i32 s4, 0x400
+; GFX940-NEXT: s_mov_b64 s[6:7], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v0
+; GFX940-NEXT: s_movk_i32 s8, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX940-NEXT: s_mov_b32 s9, 0x7060302
+; GFX940-NEXT: v_mov_b32_e32 v4, s4
+; GFX940-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX940-NEXT: v_add_f32_e32 v0, v0, v2
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v3
+; GFX940-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v0, s8
+; GFX940-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v0, v6, v7, s[4:5]
+; GFX940-NEXT: v_perm_b32 v0, v5, v0, s9
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[0:1]
+; GFX940-NEXT: buffer_atomic_cmpswap v[6:7], v4, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v6, v1
+; GFX940-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX940-NEXT: v_mov_b32_e32 v1, v6
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX940-NEXT: s_cbranch_execnz .LBB16_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_v2bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v1, s4 :: v_dual_lshlrev_b32 v2, 16, v0
+; GFX11-NEXT: s_addk_i32 s4, 0x400
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_and_b32 v3, 0xffff0000, v0
+; GFX11-NEXT: buffer_load_b32 v1, v1, s[0:3], 0 offen offset:1024
+; GFX11-NEXT: s_mov_b32 s5, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_add_f32 v5, v5, v3 :: v_dual_add_f32 v0, v0, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_add3_u32 v6, v6, v0, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v0, v6, v8, s4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v0, v5, v0, 0x7060302
+; GFX11-NEXT: v_dual_mov_b32 v6, v1 :: v_dual_mov_b32 v5, v0
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[5:6], v4, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v5
+; GFX11-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5
+; GFX11-NEXT: s_cbranch_execnz .LBB16_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_v2bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v0
+; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX10-NEXT: v_mov_b32_e32 v4, s8
+; GFX10-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: s_mov_b32 s9, 0
+; GFX10-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_add_f32_e32 v0, v0, v2
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v3
+; GFX10-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v6, v6, v0, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s8, v0, v0
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v0, v6, v8, s8
+; GFX10-NEXT: v_perm_b32 v0, v5, v0, 0x7060302
+; GFX10-NEXT: v_mov_b32_e32 v6, v1
+; GFX10-NEXT: v_mov_b32_e32 v5, v0
+; GFX10-NEXT: buffer_atomic_cmpswap v[5:6], v4, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: s_or_b32 s9, vcc_lo, s9
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s9
+; GFX10-NEXT: s_cbranch_execnz .LBB16_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s9
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_v2bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, s8
+; GFX90A-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX90A-NEXT: s_addk_i32 s8, 0x400
+; GFX90A-NEXT: s_mov_b64 s[10:11], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v0
+; GFX90A-NEXT: s_movk_i32 s12, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX90A-NEXT: s_mov_b32 s13, 0x7060302
+; GFX90A-NEXT: v_mov_b32_e32 v4, s8
+; GFX90A-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX90A-NEXT: v_add_f32_e32 v0, v0, v2
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v3
+; GFX90A-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v0, s12
+; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s12
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[8:9], v0, v0
+; GFX90A-NEXT: v_cndmask_b32_e64 v0, v6, v7, s[8:9]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v0, v5, v0, s13
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[6:7], v4, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v6, v1
+; GFX90A-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v6
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX90A-NEXT: s_cbranch_execnz .LBB16_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_v2bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, s8
+; GFX908-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_addk_i32 s8, 0x400
+; GFX908-NEXT: s_mov_b64 s[10:11], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v0
+; GFX908-NEXT: s_movk_i32 s12, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX908-NEXT: s_mov_b32 s13, 0x7060302
+; GFX908-NEXT: v_mov_b32_e32 v4, s8
+; GFX908-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX908-NEXT: v_add_f32_e32 v0, v0, v2
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v3
+; GFX908-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v6, v6, v0, s12
+; GFX908-NEXT: v_add3_u32 v8, v8, v5, s12
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[8:9], v0, v0
+; GFX908-NEXT: v_cndmask_b32_e64 v0, v6, v7, s[8:9]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v0, v5, v0, s13
+; GFX908-NEXT: v_mov_b32_e32 v6, v1
+; GFX908-NEXT: v_mov_b32_e32 v5, v0
+; GFX908-NEXT: buffer_atomic_cmpswap v[5:6], v4, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v1
+; GFX908-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX908-NEXT: s_cbranch_execnz .LBB16_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_v2bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, s8
+; GFX8-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_addk_i32 s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[10:11], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v0
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX8-NEXT: v_mov_b32_e32 v4, s8
+; GFX8-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX8-NEXT: v_add_f32_e32 v0, v0, v2
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v3
+; GFX8-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v0
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[8:9], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v6, v7, s[8:9]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v0, v5, v0, 16
+; GFX8-NEXT: v_mov_b32_e32 v6, v1
+; GFX8-NEXT: v_mov_b32_e32 v5, v0
+; GFX8-NEXT: buffer_atomic_cmpswap v[5:6], v4, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v1
+; GFX8-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX8-NEXT: v_mov_b32_e32 v1, v5
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX8-NEXT: s_cbranch_execnz .LBB16_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_v2bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dword v2, v2, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: s_addk_i32 s8, 0x400
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: s_mov_b64 s[10:11], 0
+; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v1
+; GFX7-NEXT: v_add_f32_e32 v5, v5, v0
+; GFX7-NEXT: v_alignbit_b32 v4, v3, v4, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v6
+; GFX7-NEXT: v_alignbit_b32 v3, v3, v5, 16
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v2, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
+; GFX7-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX7-NEXT: s_cbranch_execnz .LBB16_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fadd_noret_v2bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dword v2, v2, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: s_addk_i32 s8, 0x400
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: s_mov_b64 s[10:11], 0
+; GFX6-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v2
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v3
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v1
+; GFX6-NEXT: v_add_f32_e32 v5, v5, v0
+; GFX6-NEXT: v_alignbit_b32 v4, v3, v4, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v6
+; GFX6-NEXT: v_alignbit_b32 v3, v3, v5, 16
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v2, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
+; GFX6-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX6-NEXT: s_cbranch_execnz .LBB16_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(7) %ptr, i32 256
+ %unused = atomicrmw fadd ptr addrspace(7) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x bfloat> @buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset__waterfall(ptr addrspace(7) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset__waterfall:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX12-NEXT: s_mov_b32 s1, exec_lo
+; GFX12-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: buffer_load_b32 v6, v4, s[4:7], null offen offset:1024
+; GFX12-NEXT: ; implicit-def: $vgpr4
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB17_1
+; GFX12-NEXT: ; %bb.2:
+; GFX12-NEXT: s_mov_b32 exec_lo, s1
+; GFX12-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX12-NEXT: v_and_b32_e32 v9, 0xffff0000, v5
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Loop Header: Depth=1
+; GFX12-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX12-NEXT: s_mov_b32 s2, exec_lo
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_add_f32 v5, v5, v9 :: v_dual_add_f32 v4, v4, v8
+; GFX12-NEXT: v_bfe_u32 v11, v5, 16, 1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v10, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v12, 0x400000, v4
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX12-NEXT: v_or_b32_e32 v13, 0x400000, v5
+; GFX12-NEXT: v_add3_u32 v11, v11, v5, 0x7fff
+; GFX12-NEXT: v_add3_u32 v10, v10, v4, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v4, v10, v12, vcc_lo
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v11, v13, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v5, v5, v4, 0x7060302
+; GFX12-NEXT: v_mov_b32_e32 v4, v5
+; GFX12-NEXT: v_mov_b32_e32 v5, v6
+; GFX12-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX12-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB17_4
+; GFX12-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX12-NEXT: s_mov_b32 exec_lo, s2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX12-NEXT: v_mov_b32_e32 v6, v4
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB17_3
+; GFX12-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset__waterfall:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v8, 0x400, v4
+; GFX940-NEXT: s_mov_b64 s[2:3], exec
+; GFX940-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: buffer_load_dword v7, v4, s[4:7], 0 offen offset:1024
+; GFX940-NEXT: ; implicit-def: $vgpr4
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB17_1
+; GFX940-NEXT: ; %bb.2:
+; GFX940-NEXT: s_mov_b64 exec, s[2:3]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v9, 16, v5
+; GFX940-NEXT: s_movk_i32 s10, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v10, 0xffff0000, v5
+; GFX940-NEXT: s_mov_b32 s11, 0x7060302
+; GFX940-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Loop Header: Depth=1
+; GFX940-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v7
+; GFX940-NEXT: v_add_f32_e32 v4, v4, v9
+; GFX940-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX940-NEXT: v_add3_u32 v5, v5, v4, s10
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_mov_b64 s[8:9], exec
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_add_f32_e32 v5, v5, v10
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s10
+; GFX940-NEXT: v_or_b32_e32 v11, 0x400000, v5
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v11, vcc
+; GFX940-NEXT: v_perm_b32 v6, v5, v4, s11
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[6:7]
+; GFX940-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX940-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[4:7], 0 offen sc0
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB17_4
+; GFX940-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX940-NEXT: s_mov_b64 exec, s[8:9]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB17_3
+; GFX940-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset__waterfall:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: buffer_load_b32 v6, v4, s[4:7], 0 offen offset:1024
+; GFX11-NEXT: ; implicit-def: $vgpr4
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB17_1
+; GFX11-NEXT: ; %bb.2:
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v5
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Loop Header: Depth=1
+; GFX11-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_add_f32 v5, v5, v9 :: v_dual_add_f32 v4, v4, v8
+; GFX11-NEXT: v_bfe_u32 v11, v5, 16, 1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v10, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v4
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v5
+; GFX11-NEXT: v_add3_u32 v11, v11, v5, 0x7fff
+; GFX11-NEXT: v_add3_u32 v10, v10, v4, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v4, v10, v12, vcc_lo
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v11, v13, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v5, v5, v4, 0x7060302
+; GFX11-NEXT: v_mov_b32_e32 v4, v5
+; GFX11-NEXT: v_mov_b32_e32 v5, v6
+; GFX11-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX11-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], 0 offen glc
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB17_4
+; GFX11-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX11-NEXT: v_mov_b32_e32 v6, v4
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB17_3
+; GFX11-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset__waterfall:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX10-NEXT: ; implicit-def: $vgpr4
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB17_1
+; GFX10-NEXT: ; %bb.2:
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX10-NEXT: v_and_b32_e32 v9, 0xffff0000, v5
+; GFX10-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Loop Header: Depth=1
+; GFX10-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_add_f32_e32 v4, v4, v8
+; GFX10-NEXT: v_add_f32_e32 v5, v5, v9
+; GFX10-NEXT: v_bfe_u32 v10, v4, 16, 1
+; GFX10-NEXT: v_bfe_u32 v11, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v12, 0x400000, v4
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX10-NEXT: v_or_b32_e32 v13, 0x400000, v5
+; GFX10-NEXT: v_add3_u32 v10, v10, v4, 0x7fff
+; GFX10-NEXT: v_add3_u32 v11, v11, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v4, v10, v12, vcc_lo
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v11, v13, vcc_lo
+; GFX10-NEXT: v_perm_b32 v5, v5, v4, 0x7060302
+; GFX10-NEXT: v_mov_b32_e32 v4, v5
+; GFX10-NEXT: v_mov_b32_e32 v5, v6
+; GFX10-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX10-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB17_4
+; GFX10-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB17_3
+; GFX10-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset__waterfall:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v8, 0x400, v4
+; GFX90A-NEXT: s_mov_b64 s[6:7], exec
+; GFX90A-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_nop 0
+; GFX90A-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen offset:1024
+; GFX90A-NEXT: ; implicit-def: $vgpr4
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB17_1
+; GFX90A-NEXT: ; %bb.2:
+; GFX90A-NEXT: s_mov_b64 exec, s[6:7]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v9, 16, v5
+; GFX90A-NEXT: s_movk_i32 s14, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v10, 0xffff0000, v5
+; GFX90A-NEXT: s_mov_b32 s15, 0x7060302
+; GFX90A-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Loop Header: Depth=1
+; GFX90A-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v7
+; GFX90A-NEXT: v_add_f32_e32 v4, v4, v9
+; GFX90A-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX90A-NEXT: v_add3_u32 v5, v5, v4, s14
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_add_f32_e32 v5, v5, v10
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s14
+; GFX90A-NEXT: v_or_b32_e32 v11, 0x400000, v5
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v11, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v4, s15
+; GFX90A-NEXT: s_mov_b64 s[12:13], exec
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[6:7] op_sel:[0,1]
+; GFX90A-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX90A-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[8:11], 0 offen glc
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB17_4
+; GFX90A-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX90A-NEXT: s_mov_b64 exec, s[12:13]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB17_3
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset__waterfall:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v7, 0x400, v4
+; GFX908-NEXT: s_mov_b64 s[6:7], exec
+; GFX908-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_nop 0
+; GFX908-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX908-NEXT: ; implicit-def: $vgpr4
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_1
+; GFX908-NEXT: ; %bb.2:
+; GFX908-NEXT: s_mov_b64 exec, s[6:7]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX908-NEXT: s_movk_i32 s14, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v9, 0xffff0000, v5
+; GFX908-NEXT: s_mov_b32 s15, 0x7060302
+; GFX908-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Loop Header: Depth=1
+; GFX908-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX908-NEXT: v_add_f32_e32 v4, v4, v8
+; GFX908-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX908-NEXT: v_add3_u32 v5, v5, v4, s14
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v4
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX908-NEXT: v_cndmask_b32_e32 v4, v5, v10, vcc
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_add_f32_e32 v5, v5, v9
+; GFX908-NEXT: v_bfe_u32 v10, v5, 16, 1
+; GFX908-NEXT: v_add3_u32 v10, v10, v5, s14
+; GFX908-NEXT: v_or_b32_e32 v11, 0x400000, v5
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v10, v11, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v4, s15
+; GFX908-NEXT: v_mov_b32_e32 v4, v5
+; GFX908-NEXT: s_mov_b64 s[12:13], exec
+; GFX908-NEXT: v_mov_b32_e32 v5, v6
+; GFX908-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX908-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_4
+; GFX908-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX908-NEXT: s_mov_b64 exec, s[12:13]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_3
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset__waterfall:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x400, v4
+; GFX8-NEXT: s_mov_b64 s[6:7], exec
+; GFX8-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_nop 0
+; GFX8-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX8-NEXT: ; implicit-def: $vgpr4
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_1
+; GFX8-NEXT: ; %bb.2:
+; GFX8-NEXT: s_mov_b64 exec, s[6:7]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX8-NEXT: v_and_b32_e32 v9, 0xffff0000, v5
+; GFX8-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Loop Header: Depth=1
+; GFX8-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX8-NEXT: v_add_f32_e32 v4, v4, v8
+; GFX8-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v4
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v4, v5, v10, vcc
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_add_f32_e32 v5, v5, v9
+; GFX8-NEXT: v_bfe_u32 v10, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v10, vcc, v10, v5
+; GFX8-NEXT: v_add_u32_e32 v10, vcc, 0x7fff, v10
+; GFX8-NEXT: v_or_b32_e32 v11, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v10, v11, vcc
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX8-NEXT: v_mov_b32_e32 v4, v5
+; GFX8-NEXT: s_mov_b64 s[12:13], exec
+; GFX8-NEXT: v_mov_b32_e32 v5, v6
+; GFX8-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX8-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_4
+; GFX8-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX8-NEXT: s_mov_b64 exec, s[12:13]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_3
+; GFX8-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset__waterfall:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v8, vcc, 0x400, v4
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen offset:1024
+; GFX7-NEXT: ; implicit-def: $vgpr4
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_1
+; GFX7-NEXT: ; %bb.2:
+; GFX7-NEXT: s_mov_b64 exec, s[6:7]
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v5
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v6
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v7
+; GFX7-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v5
+; GFX7-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Loop Header: Depth=1
+; GFX7-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v7
+; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v7
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX7-NEXT: v_add_f32_e32 v4, v4, v10
+; GFX7-NEXT: v_add_f32_e32 v6, v6, v9
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v7
+; GFX7-NEXT: v_alignbit_b32 v5, v6, v5, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: s_mov_b64 s[12:13], exec
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX7-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v8, s[8:11], 0 offen glc
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_4
+; GFX7-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX7-NEXT: s_mov_b64 exec, s[12:13]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: v_lshlrev_b32_e32 v7, 16, v6
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_3
+; GFX7-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v0, v7
+; GFX7-NEXT: v_mov_b32_e32 v1, v4
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_v2bf16__offset__waterfall:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v8, vcc, 0x400, v4
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen offset:1024
+; GFX6-NEXT: ; implicit-def: $vgpr4
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB17_1
+; GFX6-NEXT: ; %bb.2:
+; GFX6-NEXT: s_mov_b64 exec, s[6:7]
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX6-NEXT: v_and_b32_e32 v9, 0xffff0000, v5
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v6
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v7
+; GFX6-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: v_and_b32_e32 v10, 0xffff0000, v5
+; GFX6-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Loop Header: Depth=1
+; GFX6-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v7
+; GFX6-NEXT: v_mul_f32_e32 v7, 1.0, v4
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v7
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v5
+; GFX6-NEXT: v_add_f32_e32 v4, v4, v10
+; GFX6-NEXT: v_add_f32_e32 v6, v6, v9
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v6, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v6, 16, v7
+; GFX6-NEXT: v_alignbit_b32 v5, v6, v5, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: s_mov_b64 s[12:13], exec
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX6-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v8, s[8:11], 0 offen glc
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB17_4
+; GFX6-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX6-NEXT: s_mov_b64 exec, s[12:13]
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_lshlrev_b32_e32 v7, 16, v6
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB17_3
+; GFX6-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v0, v7
+; GFX6-NEXT: v_mov_b32_e32 v1, v4
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fadd ptr addrspace(7) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+; --------------------------------------------------------------------
+; misc
+; --------------------------------------------------------------------
+
+define float @buffer_fat_ptr_system_atomic_fadd_ret_f32__offset(ptr addrspace(7) inreg %ptr, float %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_system_atomic_fadd_ret_f32__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, v0
+; GFX12-NEXT: v_mov_b32_e32 v0, s4
+; GFX12-NEXT: s_addk_co_i32 s4, 0x400
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_mov_b32_e32 v3, s4
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: buffer_load_b32 v0, v0, s[0:3], null offen offset:1024
+; GFX12-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v5, v0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v3, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB18_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_system_atomic_fadd_ret_f32__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, s4
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: buffer_atomic_add_f32 v0, v1, s[0:3], 0 offen offset:1024 sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_system_atomic_fadd_ret_f32__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, v0
+; GFX11-NEXT: v_mov_b32_e32 v0, s4
+; GFX11-NEXT: s_addk_i32 s4, 0x400
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v3, s4
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
+; GFX11-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v5, v0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v3, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB18_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_system_atomic_fadd_ret_f32__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v0
+; GFX10-NEXT: v_mov_b32_e32 v0, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: v_mov_b32_e32 v3, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v5, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB18_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_system_atomic_fadd_ret_f32__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v0
+; GFX90A-NEXT: v_mov_b32_e32 v0, s8
+; GFX90A-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX90A-NEXT: s_add_i32 s10, s8, 0x400
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_mov_b32_e32 v3, s10
+; GFX90A-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v0
+; GFX90A-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB18_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_system_atomic_fadd_ret_f32__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v0
+; GFX908-NEXT: v_mov_b32_e32 v0, s8
+; GFX908-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_add_i32 s10, s8, 0x400
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_mov_b32_e32 v3, s10
+; GFX908-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v5, v0
+; GFX908-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB18_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_system_atomic_fadd_ret_f32__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v0
+; GFX8-NEXT: v_mov_b32_e32 v0, s8
+; GFX8-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_add_i32 s10, s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_mov_b32_e32 v3, s10
+; GFX8-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v5, v0
+; GFX8-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: v_mov_b32_e32 v1, v5
+; GFX8-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB18_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_system_atomic_fadd_ret_f32__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v0
+; GFX7-NEXT: v_mov_b32_e32 v0, s8
+; GFX7-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: s_add_i32 s10, s8, 0x400
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mov_b32_e32 v3, s10
+; GFX7-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v0
+; GFX7-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX7-NEXT: v_mov_b32_e32 v0, v4
+; GFX7-NEXT: v_mov_b32_e32 v1, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB18_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_system_atomic_fadd_ret_f32__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v0
+; GFX6-NEXT: v_mov_b32_e32 v0, s8
+; GFX6-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: s_add_i32 s10, s8, 0x400
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mov_b32_e32 v3, s10
+; GFX6-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v0
+; GFX6-NEXT: v_add_f32_e32 v4, v5, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v0, v4
+; GFX6-NEXT: v_mov_b32_e32 v1, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB18_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fadd ptr addrspace(7) %gep, float %val seq_cst
+ ret float %result
+}
+
+attributes #0 = { nounwind "amdgpu-unsafe-fp-atomics"="true" }
+
+
diff --git a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmax.ll b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmax.ll
new file mode 100644
index 0000000000000..684fa2d7df60a
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmax.ll
@@ -0,0 +1,8062 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 < %s | FileCheck -check-prefix=GFX12 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx940 < %s | FileCheck -check-prefix=GFX940 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx90a < %s | FileCheck -check-prefix=GFX90A %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx908 < %s | FileCheck -check-prefix=GFX908 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tonga < %s | FileCheck -check-prefix=GFX8 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=hawaii < %s | FileCheck -check-prefix=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tahiti < %s | FileCheck -check-prefix=GFX6 %s
+
+; --------------------------------------------------------------------
+; float
+; --------------------------------------------------------------------
+
+define float @buffer_fat_ptr_agent_atomic_fmax_ret_f32__offset(ptr addrspace(7) inreg %ptr, float %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f32__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s4
+; GFX12-NEXT: s_addk_co_i32 s4, 0x400
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: v_dual_mov_b32 v3, s4 :: v_dual_max_num_f32 v2, v1, v1
+; GFX12-NEXT: buffer_load_b32 v0, v0, s[0:3], null offen offset:1024
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v5, v0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v0, v5, v5
+; GFX12-NEXT: v_max_num_f32_e32 v4, v0, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v3, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB0_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f32__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, v0
+; GFX940-NEXT: v_mov_b32_e32 v0, s4
+; GFX940-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen offset:1024
+; GFX940-NEXT: s_add_i32 s6, s4, 0x400
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX940-NEXT: v_mov_b32_e32 v3, s6
+; GFX940-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v0
+; GFX940-NEXT: v_max_f32_e32 v0, v5, v5
+; GFX940-NEXT: v_max_f32_e32 v4, v0, v2
+; GFX940-NEXT: v_mov_b64_e32 v[0:1], v[4:5]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB0_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f32__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s4
+; GFX11-NEXT: s_addk_i32 s4, 0x400
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: v_dual_mov_b32 v3, s4 :: v_dual_max_f32 v2, v1, v1
+; GFX11-NEXT: buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v5, v0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v0, v5, v5
+; GFX11-NEXT: v_max_f32_e32 v4, v0, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v3, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB0_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f32__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_mov_b32_e32 v0, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: v_mov_b32_e32 v3, s8
+; GFX10-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX10-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v5, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_max_f32_e32 v0, v5, v5
+; GFX10-NEXT: v_max_f32_e32 v4, v0, v2
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB0_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f32__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: v_mov_b32_e32 v0, s8
+; GFX90A-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX90A-NEXT: s_add_i32 s10, s8, 0x400
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX90A-NEXT: v_mov_b32_e32 v3, s10
+; GFX90A-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v0
+; GFX90A-NEXT: v_max_f32_e32 v0, v5, v5
+; GFX90A-NEXT: v_max_f32_e32 v4, v0, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB0_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f32__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: v_mov_b32_e32 v0, s8
+; GFX908-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_add_i32 s10, s8, 0x400
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX908-NEXT: v_mov_b32_e32 v3, s10
+; GFX908-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v5, v0
+; GFX908-NEXT: v_max_f32_e32 v0, v5, v5
+; GFX908-NEXT: v_max_f32_e32 v4, v0, v2
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB0_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f32__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_mov_b32_e32 v0, s8
+; GFX8-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_add_i32 s10, s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX8-NEXT: v_mov_b32_e32 v3, s10
+; GFX8-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v5, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v5
+; GFX8-NEXT: v_max_f32_e32 v4, v0, v2
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: v_mov_b32_e32 v1, v5
+; GFX8-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB0_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f32__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v1, v0
+; GFX7-NEXT: v_mov_b32_e32 v0, s8
+; GFX7-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: s_add_i32 s10, s8, 0x400
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, s10
+; GFX7-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v5
+; GFX7-NEXT: v_max_f32_e32 v4, v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v0, v4
+; GFX7-NEXT: v_mov_b32_e32 v1, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB0_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f32__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v1, v0
+; GFX6-NEXT: v_mov_b32_e32 v0, s8
+; GFX6-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: s_add_i32 s10, s8, 0x400
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, s10
+; GFX6-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v5
+; GFX6-NEXT: v_max_f32_e32 v4, v0, v2
+; GFX6-NEXT: v_mov_b32_e32 v0, v4
+; GFX6-NEXT: v_mov_b32_e32 v1, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB0_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmax ptr addrspace(7) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define void @buffer_fat_ptr_agent_atomic_fmax_noret_f32__offset(ptr addrspace(7) inreg %ptr, float %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f32__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v1, s4 :: v_dual_max_num_f32 v2, v0, v0
+; GFX12-NEXT: s_addk_co_i32 s4, 0x400
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_mov_b32_e32 v3, s4
+; GFX12-NEXT: buffer_load_b32 v1, v1, s[0:3], null offen offset:1024
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v0, v1, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v0, v0, v2
+; GFX12-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v4
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB1_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f32__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, s4
+; GFX940-NEXT: buffer_load_dword v1, v1, s[0:3], 0 offen offset:1024
+; GFX940-NEXT: s_add_i32 s6, s4, 0x400
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v0, v0
+; GFX940-NEXT: v_mov_b32_e32 v3, s6
+; GFX940-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v0, v1, v1
+; GFX940-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[0:1]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB1_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f32__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v1, s4 :: v_dual_max_f32 v2, v0, v0
+; GFX11-NEXT: s_addk_i32 s4, 0x400
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v3, s4
+; GFX11-NEXT: buffer_load_b32 v1, v1, s[0:3], 0 offen offset:1024
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v0, v1, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX11-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v4
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB1_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f32__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: v_max_f32_e32 v2, v0, v0
+; GFX10-NEXT: v_mov_b32_e32 v3, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v0, v1, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX10-NEXT: v_mov_b32_e32 v5, v1
+; GFX10-NEXT: v_mov_b32_e32 v4, v0
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v4
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB1_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f32__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, s8
+; GFX90A-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX90A-NEXT: s_add_i32 s10, s8, 0x400
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v0, v0
+; GFX90A-NEXT: v_mov_b32_e32 v3, s10
+; GFX90A-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v0, v1, v1
+; GFX90A-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB1_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f32__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, s8
+; GFX908-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_add_i32 s10, s8, 0x400
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v0, v0
+; GFX908-NEXT: v_mov_b32_e32 v3, s10
+; GFX908-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v0, v1, v1
+; GFX908-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX908-NEXT: v_mov_b32_e32 v5, v1
+; GFX908-NEXT: v_mov_b32_e32 v4, v0
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: v_mov_b32_e32 v1, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB1_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f32__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, s8
+; GFX8-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_add_i32 s10, s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v0
+; GFX8-NEXT: v_mov_b32_e32 v3, s10
+; GFX8-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v1
+; GFX8-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX8-NEXT: v_mov_b32_e32 v5, v1
+; GFX8-NEXT: v_mov_b32_e32 v4, v0
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v1, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB1_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f32__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v1, s8
+; GFX7-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: s_add_i32 s10, s8, 0x400
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v0
+; GFX7-NEXT: v_mov_b32_e32 v3, s10
+; GFX7-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v1
+; GFX7-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v1
+; GFX7-NEXT: v_mov_b32_e32 v4, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v1, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB1_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f32__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v1, s8
+; GFX6-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: s_add_i32 s10, s8, 0x400
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v0
+; GFX6-NEXT: v_mov_b32_e32 v3, s10
+; GFX6-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v1
+; GFX6-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v1
+; GFX6-NEXT: v_mov_b32_e32 v4, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v1, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB1_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(7) %ptr, i32 256
+ %unused = atomicrmw fmax ptr addrspace(7) %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define float @buffer_fat_ptr_agent_atomic_fmax_ret_f32__offset__waterfall(ptr addrspace(7) %ptr, float %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f32__offset__waterfall:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX12-NEXT: s_mov_b32 s1, exec_lo
+; GFX12-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: buffer_load_b32 v6, v4, s[4:7], null offen offset:1024
+; GFX12-NEXT: ; implicit-def: $vgpr4
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB2_1
+; GFX12-NEXT: ; %bb.2:
+; GFX12-NEXT: s_mov_b32 exec_lo, s1
+; GFX12-NEXT: v_max_num_f32_e32 v8, v5, v5
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Loop Header: Depth=1
+; GFX12-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v4, v6, v6
+; GFX12-NEXT: s_mov_b32 s2, exec_lo
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v5, v4, v8
+; GFX12-NEXT: v_mov_b32_e32 v4, v5
+; GFX12-NEXT: v_mov_b32_e32 v5, v6
+; GFX12-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX12-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB2_4
+; GFX12-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX12-NEXT: s_mov_b32 exec_lo, s2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX12-NEXT: v_mov_b32_e32 v6, v4
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB2_3
+; GFX12-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f32__offset__waterfall:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v8, 0x400, v4
+; GFX940-NEXT: s_mov_b64 s[2:3], exec
+; GFX940-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: buffer_load_dword v7, v4, s[4:7], 0 offen offset:1024
+; GFX940-NEXT: ; implicit-def: $vgpr4
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB2_1
+; GFX940-NEXT: ; %bb.2:
+; GFX940-NEXT: s_mov_b64 exec, s[2:3]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_max_f32_e32 v9, v5, v5
+; GFX940-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Loop Header: Depth=1
+; GFX940-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v4, v7, v7
+; GFX940-NEXT: v_max_f32_e32 v6, v4, v9
+; GFX940-NEXT: s_mov_b64 s[8:9], exec
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[6:7]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX940-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[4:7], 0 offen sc0
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB2_4
+; GFX940-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX940-NEXT: s_mov_b64 exec, s[8:9]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB2_3
+; GFX940-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f32__offset__waterfall:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: buffer_load_b32 v6, v4, s[4:7], 0 offen offset:1024
+; GFX11-NEXT: ; implicit-def: $vgpr4
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB2_1
+; GFX11-NEXT: ; %bb.2:
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: v_max_f32_e32 v8, v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Loop Header: Depth=1
+; GFX11-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v4, v6, v6
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v5, v4, v8
+; GFX11-NEXT: v_mov_b32_e32 v4, v5
+; GFX11-NEXT: v_mov_b32_e32 v5, v6
+; GFX11-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX11-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], 0 offen glc
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB2_4
+; GFX11-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX11-NEXT: v_mov_b32_e32 v6, v4
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB2_3
+; GFX11-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f32__offset__waterfall:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX10-NEXT: ; implicit-def: $vgpr4
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB2_1
+; GFX10-NEXT: ; %bb.2:
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: v_max_f32_e32 v8, v5, v5
+; GFX10-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Loop Header: Depth=1
+; GFX10-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v4, v6, v6
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_max_f32_e32 v5, v4, v8
+; GFX10-NEXT: v_mov_b32_e32 v4, v5
+; GFX10-NEXT: v_mov_b32_e32 v5, v6
+; GFX10-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX10-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB2_4
+; GFX10-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB2_3
+; GFX10-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f32__offset__waterfall:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v8, 0x400, v4
+; GFX90A-NEXT: s_mov_b64 s[6:7], exec
+; GFX90A-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_nop 0
+; GFX90A-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen offset:1024
+; GFX90A-NEXT: ; implicit-def: $vgpr4
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB2_1
+; GFX90A-NEXT: ; %bb.2:
+; GFX90A-NEXT: s_mov_b64 exec, s[6:7]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_max_f32_e32 v9, v5, v5
+; GFX90A-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Loop Header: Depth=1
+; GFX90A-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v4, v7, v7
+; GFX90A-NEXT: v_max_f32_e32 v6, v4, v9
+; GFX90A-NEXT: s_mov_b64 s[12:13], exec
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[6:7] op_sel:[0,1]
+; GFX90A-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX90A-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[8:11], 0 offen glc
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB2_4
+; GFX90A-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX90A-NEXT: s_mov_b64 exec, s[12:13]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB2_3
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f32__offset__waterfall:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v7, 0x400, v4
+; GFX908-NEXT: s_mov_b64 s[6:7], exec
+; GFX908-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_nop 0
+; GFX908-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX908-NEXT: ; implicit-def: $vgpr4
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB2_1
+; GFX908-NEXT: ; %bb.2:
+; GFX908-NEXT: s_mov_b64 exec, s[6:7]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_max_f32_e32 v8, v5, v5
+; GFX908-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Loop Header: Depth=1
+; GFX908-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v4, v6, v6
+; GFX908-NEXT: v_max_f32_e32 v5, v4, v8
+; GFX908-NEXT: v_mov_b32_e32 v4, v5
+; GFX908-NEXT: s_mov_b64 s[12:13], exec
+; GFX908-NEXT: v_mov_b32_e32 v5, v6
+; GFX908-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX908-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB2_4
+; GFX908-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX908-NEXT: s_mov_b64 exec, s[12:13]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB2_3
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f32__offset__waterfall:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x400, v4
+; GFX8-NEXT: s_mov_b64 s[6:7], exec
+; GFX8-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_nop 0
+; GFX8-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX8-NEXT: ; implicit-def: $vgpr4
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB2_1
+; GFX8-NEXT: ; %bb.2:
+; GFX8-NEXT: s_mov_b64 exec, s[6:7]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_mul_f32_e32 v8, 1.0, v5
+; GFX8-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Loop Header: Depth=1
+; GFX8-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v6
+; GFX8-NEXT: v_max_f32_e32 v5, v4, v8
+; GFX8-NEXT: v_mov_b32_e32 v4, v5
+; GFX8-NEXT: s_mov_b64 s[12:13], exec
+; GFX8-NEXT: v_mov_b32_e32 v5, v6
+; GFX8-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX8-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB2_4
+; GFX8-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX8-NEXT: s_mov_b64 exec, s[12:13]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB2_3
+; GFX8-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f32__offset__waterfall:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v7, vcc, 0x400, v4
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX7-NEXT: ; implicit-def: $vgpr4
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_1
+; GFX7-NEXT: ; %bb.2:
+; GFX7-NEXT: s_mov_b64 exec, s[6:7]
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v5
+; GFX7-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Loop Header: Depth=1
+; GFX7-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v6
+; GFX7-NEXT: v_max_f32_e32 v5, v4, v8
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_mov_b64 s[12:13], exec
+; GFX7-NEXT: v_mov_b32_e32 v5, v6
+; GFX7-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX7-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_4
+; GFX7-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX7-NEXT: s_mov_b64 exec, s[12:13]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_3
+; GFX7-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v0, v4
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f32__offset__waterfall:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v7, vcc, 0x400, v4
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX6-NEXT: ; implicit-def: $vgpr4
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB2_1
+; GFX6-NEXT: ; %bb.2:
+; GFX6-NEXT: s_mov_b64 exec, s[6:7]
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: v_mul_f32_e32 v8, 1.0, v5
+; GFX6-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Loop Header: Depth=1
+; GFX6-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v6
+; GFX6-NEXT: v_max_f32_e32 v5, v4, v8
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_mov_b64 s[12:13], exec
+; GFX6-NEXT: v_mov_b32_e32 v5, v6
+; GFX6-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX6-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB2_4
+; GFX6-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX6-NEXT: s_mov_b64 exec, s[12:13]
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB2_3
+; GFX6-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v0, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmax ptr addrspace(7) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+; --------------------------------------------------------------------
+; double
+; --------------------------------------------------------------------
+
+define double @buffer_fat_ptr_agent_atomic_fmax_ret_f64__offset(ptr addrspace(7) inreg %ptr, double %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f64__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX12-NEXT: v_mov_b32_e32 v0, s4
+; GFX12-NEXT: s_addk_co_i32 s4, 0x800
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_mov_b32_e32 v6, s4
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[2:3], v[2:3]
+; GFX12-NEXT: buffer_load_b64 v[0:1], v0, s[0:3], null offen offset:2048
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v10, v1 :: v_dual_mov_b32 v9, v0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[0:1], v[9:10], v[9:10]
+; GFX12-NEXT: v_max_num_f64_e32 v[7:8], v[0:1], v[4:5]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v0, v7 :: v_dual_mov_b32 v1, v8
+; GFX12-NEXT: v_dual_mov_b32 v2, v9 :: v_dual_mov_b32 v3, v10
+; GFX12-NEXT: buffer_atomic_cmpswap_b64 v[0:3], v6, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[9:10]
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB3_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f64__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v0
+; GFX940-NEXT: v_mov_b32_e32 v0, s4
+; GFX940-NEXT: v_mov_b32_e32 v3, v1
+; GFX940-NEXT: buffer_load_dwordx2 v[0:1], v0, s[0:3], 0 offen offset:2048
+; GFX940-NEXT: s_add_i32 s6, s4, 0x800
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v6, s6
+; GFX940-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[10:11], v[0:1]
+; GFX940-NEXT: v_max_f64 v[0:1], v[10:11], v[10:11]
+; GFX940-NEXT: v_max_f64 v[8:9], v[0:1], v[4:5]
+; GFX940-NEXT: v_mov_b64_e32 v[0:1], v[8:9]
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], v[10:11]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v6, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB3_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f64__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX11-NEXT: v_mov_b32_e32 v0, s4
+; GFX11-NEXT: s_addk_i32 s4, 0x800
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_mov_b32_e32 v6, s4
+; GFX11-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX11-NEXT: buffer_load_b64 v[0:1], v0, s[0:3], 0 offen offset:2048
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v10, v1 :: v_dual_mov_b32 v9, v0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[0:1], v[9:10], v[9:10]
+; GFX11-NEXT: v_max_f64 v[7:8], v[0:1], v[4:5]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v0, v7 :: v_dual_mov_b32 v1, v8
+; GFX11-NEXT: v_dual_mov_b32 v2, v9 :: v_dual_mov_b32 v3, v10
+; GFX11-NEXT: buffer_atomic_cmpswap_b64 v[0:3], v6, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[9:10]
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB3_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f64__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v0
+; GFX10-NEXT: v_mov_b32_e32 v0, s8
+; GFX10-NEXT: v_mov_b32_e32 v3, v1
+; GFX10-NEXT: s_addk_i32 s8, 0x800
+; GFX10-NEXT: v_mov_b32_e32 v6, s8
+; GFX10-NEXT: buffer_load_dwordx2 v[0:1], v0, s[4:7], 0 offen offset:2048
+; GFX10-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v10, v1
+; GFX10-NEXT: v_mov_b32_e32 v9, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_max_f64 v[0:1], v[9:10], v[9:10]
+; GFX10-NEXT: v_max_f64 v[7:8], v[0:1], v[4:5]
+; GFX10-NEXT: v_mov_b32_e32 v0, v7
+; GFX10-NEXT: v_mov_b32_e32 v1, v8
+; GFX10-NEXT: v_mov_b32_e32 v2, v9
+; GFX10-NEXT: v_mov_b32_e32 v3, v10
+; GFX10-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v6, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[9:10]
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB3_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f64__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v0
+; GFX90A-NEXT: v_mov_b32_e32 v0, s8
+; GFX90A-NEXT: v_mov_b32_e32 v3, v1
+; GFX90A-NEXT: buffer_load_dwordx2 v[0:1], v0, s[4:7], 0 offen offset:2048
+; GFX90A-NEXT: s_add_i32 s10, s8, 0x800
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX90A-NEXT: v_mov_b32_e32 v6, s10
+; GFX90A-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[10:11], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: v_max_f64 v[0:1], v[10:11], v[10:11]
+; GFX90A-NEXT: v_max_f64 v[8:9], v[0:1], v[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[8:9], v[8:9] op_sel:[0,1]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[10:11], v[10:11] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v6, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB3_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f64__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v0
+; GFX908-NEXT: v_mov_b32_e32 v0, s8
+; GFX908-NEXT: v_mov_b32_e32 v3, v1
+; GFX908-NEXT: buffer_load_dwordx2 v[0:1], v0, s[4:7], 0 offen offset:2048
+; GFX908-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX908-NEXT: s_add_i32 s10, s8, 0x800
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_mov_b32_e32 v6, s10
+; GFX908-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v10, v1
+; GFX908-NEXT: v_mov_b32_e32 v9, v0
+; GFX908-NEXT: v_max_f64 v[0:1], v[9:10], v[9:10]
+; GFX908-NEXT: v_max_f64 v[7:8], v[0:1], v[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v7
+; GFX908-NEXT: v_mov_b32_e32 v1, v8
+; GFX908-NEXT: v_mov_b32_e32 v2, v9
+; GFX908-NEXT: v_mov_b32_e32 v3, v10
+; GFX908-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v6, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[9:10]
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB3_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f64__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v0
+; GFX8-NEXT: v_mov_b32_e32 v0, s8
+; GFX8-NEXT: v_mov_b32_e32 v3, v1
+; GFX8-NEXT: buffer_load_dwordx2 v[0:1], v0, s[4:7], 0 offen offset:2048
+; GFX8-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX8-NEXT: s_add_i32 s10, s8, 0x800
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_mov_b32_e32 v6, s10
+; GFX8-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v10, v1
+; GFX8-NEXT: v_mov_b32_e32 v9, v0
+; GFX8-NEXT: v_max_f64 v[0:1], v[9:10], v[9:10]
+; GFX8-NEXT: v_max_f64 v[7:8], v[0:1], v[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v7
+; GFX8-NEXT: v_mov_b32_e32 v1, v8
+; GFX8-NEXT: v_mov_b32_e32 v2, v9
+; GFX8-NEXT: v_mov_b32_e32 v3, v10
+; GFX8-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v6, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[9:10]
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB3_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f64__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v0
+; GFX7-NEXT: v_mov_b32_e32 v0, s8
+; GFX7-NEXT: v_mov_b32_e32 v3, v1
+; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v0, s[4:7], 0 offen offset:2048
+; GFX7-NEXT: s_add_i32 s10, s8, 0x800
+; GFX7-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mov_b32_e32 v6, s10
+; GFX7-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v10, v1
+; GFX7-NEXT: v_mov_b32_e32 v9, v0
+; GFX7-NEXT: v_max_f64 v[0:1], v[9:10], v[9:10]
+; GFX7-NEXT: v_max_f64 v[7:8], v[0:1], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v7
+; GFX7-NEXT: v_mov_b32_e32 v1, v8
+; GFX7-NEXT: v_mov_b32_e32 v2, v9
+; GFX7-NEXT: v_mov_b32_e32 v3, v10
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v6, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[9:10]
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB3_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f64__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v0
+; GFX6-NEXT: v_mov_b32_e32 v0, s8
+; GFX6-NEXT: v_mov_b32_e32 v3, v1
+; GFX6-NEXT: buffer_load_dwordx2 v[0:1], v0, s[4:7], 0 offen offset:2048
+; GFX6-NEXT: s_add_i32 s10, s8, 0x800
+; GFX6-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mov_b32_e32 v6, s10
+; GFX6-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v10, v1
+; GFX6-NEXT: v_mov_b32_e32 v9, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_max_f64 v[0:1], v[9:10], v[9:10]
+; GFX6-NEXT: v_max_f64 v[7:8], v[0:1], v[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v0, v7
+; GFX6-NEXT: v_mov_b32_e32 v1, v8
+; GFX6-NEXT: v_mov_b32_e32 v2, v9
+; GFX6-NEXT: v_mov_b32_e32 v3, v10
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v6, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[9:10]
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB3_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmax ptr addrspace(7) %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define void @buffer_fat_ptr_agent_atomic_fmax_noret_f64__offset(ptr addrspace(7) inreg %ptr, double %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f64__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, s4
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[0:1], v[0:1]
+; GFX12-NEXT: s_addk_co_i32 s4, 0x800
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_mov_b32_e32 v6, s4
+; GFX12-NEXT: buffer_load_b64 v[2:3], v2, s[0:3], null offen offset:2048
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f64_e32 v[0:1], v[2:3], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_max_num_f64_e32 v[0:1], v[0:1], v[4:5]
+; GFX12-NEXT: v_dual_mov_b32 v10, v3 :: v_dual_mov_b32 v9, v2
+; GFX12-NEXT: v_dual_mov_b32 v8, v1 :: v_dual_mov_b32 v7, v0
+; GFX12-NEXT: buffer_atomic_cmpswap_b64 v[7:10], v6, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[7:8], v[2:3]
+; GFX12-NEXT: v_dual_mov_b32 v2, v7 :: v_dual_mov_b32 v3, v8
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB4_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f64__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, s4
+; GFX940-NEXT: buffer_load_dwordx2 v[2:3], v2, s[0:3], 0 offen offset:2048
+; GFX940-NEXT: s_add_i32 s6, s4, 0x800
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v6, s6
+; GFX940-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX940-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
+; GFX940-NEXT: v_mov_b64_e32 v[10:11], v[2:3]
+; GFX940-NEXT: v_mov_b64_e32 v[8:9], v[0:1]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v6, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[2:3]
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], v[8:9]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB4_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f64__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, s4
+; GFX11-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
+; GFX11-NEXT: s_addk_i32 s4, 0x800
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v6, s4
+; GFX11-NEXT: buffer_load_b64 v[2:3], v2, s[0:3], 0 offen offset:2048
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
+; GFX11-NEXT: v_dual_mov_b32 v10, v3 :: v_dual_mov_b32 v9, v2
+; GFX11-NEXT: v_dual_mov_b32 v8, v1 :: v_dual_mov_b32 v7, v0
+; GFX11-NEXT: buffer_atomic_cmpswap_b64 v[7:10], v6, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[7:8], v[2:3]
+; GFX11-NEXT: v_dual_mov_b32 v2, v7 :: v_dual_mov_b32 v3, v8
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB4_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f64__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, s8
+; GFX10-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
+; GFX10-NEXT: s_addk_i32 s8, 0x800
+; GFX10-NEXT: v_mov_b32_e32 v6, s8
+; GFX10-NEXT: buffer_load_dwordx2 v[2:3], v2, s[4:7], 0 offen offset:2048
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
+; GFX10-NEXT: v_mov_b32_e32 v10, v3
+; GFX10-NEXT: v_mov_b32_e32 v9, v2
+; GFX10-NEXT: v_mov_b32_e32 v8, v1
+; GFX10-NEXT: v_mov_b32_e32 v7, v0
+; GFX10-NEXT: buffer_atomic_cmpswap_x2 v[7:10], v6, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[7:8], v[2:3]
+; GFX10-NEXT: v_mov_b32_e32 v2, v7
+; GFX10-NEXT: v_mov_b32_e32 v3, v8
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB4_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f64__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, s8
+; GFX90A-NEXT: buffer_load_dwordx2 v[2:3], v2, s[4:7], 0 offen offset:2048
+; GFX90A-NEXT: s_add_i32 s10, s8, 0x800
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
+; GFX90A-NEXT: v_mov_b32_e32 v6, s10
+; GFX90A-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX90A-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[10:11], v[2:3], v[2:3] op_sel:[0,1]
+; GFX90A-NEXT: v_pk_mov_b32 v[8:9], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v6, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[2:3]
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[8:9], v[8:9] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB4_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f64__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, s8
+; GFX908-NEXT: buffer_load_dwordx2 v[2:3], v2, s[4:7], 0 offen offset:2048
+; GFX908-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
+; GFX908-NEXT: s_add_i32 s10, s8, 0x800
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_mov_b32_e32 v6, s10
+; GFX908-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX908-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v10, v3
+; GFX908-NEXT: v_mov_b32_e32 v9, v2
+; GFX908-NEXT: v_mov_b32_e32 v8, v1
+; GFX908-NEXT: v_mov_b32_e32 v7, v0
+; GFX908-NEXT: buffer_atomic_cmpswap_x2 v[7:10], v6, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[7:8], v[2:3]
+; GFX908-NEXT: v_mov_b32_e32 v2, v7
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: v_mov_b32_e32 v3, v8
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB4_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f64__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, s8
+; GFX8-NEXT: buffer_load_dwordx2 v[2:3], v2, s[4:7], 0 offen offset:2048
+; GFX8-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
+; GFX8-NEXT: s_add_i32 s10, s8, 0x800
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_mov_b32_e32 v6, s10
+; GFX8-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX8-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v10, v3
+; GFX8-NEXT: v_mov_b32_e32 v9, v2
+; GFX8-NEXT: v_mov_b32_e32 v8, v1
+; GFX8-NEXT: v_mov_b32_e32 v7, v0
+; GFX8-NEXT: buffer_atomic_cmpswap_x2 v[7:10], v6, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[7:8], v[2:3]
+; GFX8-NEXT: v_mov_b32_e32 v2, v7
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v3, v8
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB4_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f64__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dwordx2 v[2:3], v2, s[4:7], 0 offen offset:2048
+; GFX7-NEXT: s_add_i32 s10, s8, 0x800
+; GFX7-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mov_b32_e32 v6, s10
+; GFX7-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX7-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v10, v3
+; GFX7-NEXT: v_mov_b32_e32 v9, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v1
+; GFX7-NEXT: v_mov_b32_e32 v7, v0
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[7:10], v6, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[7:8], v[2:3]
+; GFX7-NEXT: v_mov_b32_e32 v2, v7
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v8
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB4_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f64__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dwordx2 v[2:3], v2, s[4:7], 0 offen offset:2048
+; GFX6-NEXT: s_add_i32 s10, s8, 0x800
+; GFX6-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mov_b32_e32 v6, s10
+; GFX6-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX6-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v10, v3
+; GFX6-NEXT: v_mov_b32_e32 v9, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v1
+; GFX6-NEXT: v_mov_b32_e32 v7, v0
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[7:10], v6, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[7:8], v[2:3]
+; GFX6-NEXT: v_mov_b32_e32 v2, v7
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v8
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB4_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(7) %ptr, i32 256
+ %unused = atomicrmw fmax ptr addrspace(7) %gep, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+define double @buffer_fat_ptr_agent_atomic_fmax_ret_f64__offset__waterfall(ptr addrspace(7) %ptr, double %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f64__offset__waterfall:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v8, v3 :: v_dual_mov_b32 v7, v2
+; GFX12-NEXT: v_dual_mov_b32 v10, v1 :: v_dual_mov_b32 v9, v0
+; GFX12-NEXT: v_add_nc_u32_e32 v15, 0x800, v4
+; GFX12-NEXT: s_mov_b32 s1, exec_lo
+; GFX12-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_readfirstlane_b32 s4, v9
+; GFX12-NEXT: v_readfirstlane_b32 s5, v10
+; GFX12-NEXT: v_readfirstlane_b32 s6, v7
+; GFX12-NEXT: v_readfirstlane_b32 s7, v8
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[9:10]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[7:8]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: buffer_load_b64 v[13:14], v4, s[4:7], null offen offset:2048
+; GFX12-NEXT: ; implicit-def: $vgpr4
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB5_1
+; GFX12-NEXT: ; %bb.2:
+; GFX12-NEXT: s_mov_b32 exec_lo, s1
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[5:6], v[5:6]
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Loop Header: Depth=1
+; GFX12-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f64_e32 v[0:1], v[13:14], v[13:14]
+; GFX12-NEXT: s_mov_b32 s2, exec_lo
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[11:12], v[0:1], v[4:5]
+; GFX12-NEXT: v_dual_mov_b32 v0, v11 :: v_dual_mov_b32 v1, v12
+; GFX12-NEXT: v_dual_mov_b32 v2, v13 :: v_dual_mov_b32 v3, v14
+; GFX12-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX12-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX12-NEXT: v_readfirstlane_b32 s4, v9
+; GFX12-NEXT: v_readfirstlane_b32 s5, v10
+; GFX12-NEXT: v_readfirstlane_b32 s6, v7
+; GFX12-NEXT: v_readfirstlane_b32 s7, v8
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[9:10]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[7:8]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: buffer_atomic_cmpswap_b64 v[0:3], v15, s[4:7], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB5_4
+; GFX12-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX12-NEXT: s_mov_b32 exec_lo, s2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[13:14]
+; GFX12-NEXT: v_dual_mov_b32 v14, v1 :: v_dual_mov_b32 v13, v0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB5_3
+; GFX12-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f64__offset__waterfall:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v11, v6
+; GFX940-NEXT: v_mov_b32_e32 v10, v5
+; GFX940-NEXT: v_mov_b32_e32 v9, v3
+; GFX940-NEXT: v_mov_b32_e32 v8, v2
+; GFX940-NEXT: v_mov_b32_e32 v7, v1
+; GFX940-NEXT: v_mov_b32_e32 v6, v0
+; GFX940-NEXT: v_add_u32_e32 v14, 0x800, v4
+; GFX940-NEXT: s_mov_b64 s[2:3], exec
+; GFX940-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: v_readfirstlane_b32 s4, v6
+; GFX940-NEXT: v_readfirstlane_b32 s5, v7
+; GFX940-NEXT: v_readfirstlane_b32 s6, v8
+; GFX940-NEXT: v_readfirstlane_b32 s7, v9
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[6:7]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[8:9]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: buffer_load_dwordx2 v[12:13], v4, s[4:7], 0 offen offset:2048
+; GFX940-NEXT: ; implicit-def: $vgpr4
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB5_1
+; GFX940-NEXT: ; %bb.2:
+; GFX940-NEXT: s_mov_b64 exec, s[2:3]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_max_f64 v[4:5], v[10:11], v[10:11]
+; GFX940-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Loop Header: Depth=1
+; GFX940-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f64 v[0:1], v[12:13], v[12:13]
+; GFX940-NEXT: v_max_f64 v[10:11], v[0:1], v[4:5]
+; GFX940-NEXT: v_mov_b64_e32 v[0:1], v[10:11]
+; GFX940-NEXT: s_mov_b64 s[8:9], exec
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], v[12:13]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX940-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX940-NEXT: v_readfirstlane_b32 s4, v6
+; GFX940-NEXT: v_readfirstlane_b32 s5, v7
+; GFX940-NEXT: v_readfirstlane_b32 s6, v8
+; GFX940-NEXT: v_readfirstlane_b32 s7, v9
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[6:7]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[8:9]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v14, s[4:7], 0 offen sc0
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB5_4
+; GFX940-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX940-NEXT: s_mov_b64 exec, s[8:9]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[12:13]
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b64_e32 v[12:13], v[0:1]
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB5_3
+; GFX940-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f64__offset__waterfall:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v8, v3 :: v_dual_mov_b32 v7, v2
+; GFX11-NEXT: v_dual_mov_b32 v10, v1 :: v_dual_mov_b32 v9, v0
+; GFX11-NEXT: v_add_nc_u32_e32 v15, 0x800, v4
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_readfirstlane_b32 s4, v9
+; GFX11-NEXT: v_readfirstlane_b32 s5, v10
+; GFX11-NEXT: v_readfirstlane_b32 s6, v7
+; GFX11-NEXT: v_readfirstlane_b32 s7, v8
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[9:10]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[7:8]
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: buffer_load_b64 v[13:14], v4, s[4:7], 0 offen offset:2048
+; GFX11-NEXT: ; implicit-def: $vgpr4
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB5_1
+; GFX11-NEXT: ; %bb.2:
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: v_max_f64 v[4:5], v[5:6], v[5:6]
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Loop Header: Depth=1
+; GFX11-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f64 v[0:1], v[13:14], v[13:14]
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[11:12], v[0:1], v[4:5]
+; GFX11-NEXT: v_dual_mov_b32 v0, v11 :: v_dual_mov_b32 v1, v12
+; GFX11-NEXT: v_dual_mov_b32 v2, v13 :: v_dual_mov_b32 v3, v14
+; GFX11-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX11-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX11-NEXT: v_readfirstlane_b32 s4, v9
+; GFX11-NEXT: v_readfirstlane_b32 s5, v10
+; GFX11-NEXT: v_readfirstlane_b32 s6, v7
+; GFX11-NEXT: v_readfirstlane_b32 s7, v8
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[9:10]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[7:8]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_atomic_cmpswap_b64 v[0:3], v15, s[4:7], 0 offen glc
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB5_4
+; GFX11-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[13:14]
+; GFX11-NEXT: v_dual_mov_b32 v14, v1 :: v_dual_mov_b32 v13, v0
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB5_3
+; GFX11-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f64__offset__waterfall:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v8, v3
+; GFX10-NEXT: v_mov_b32_e32 v7, v2
+; GFX10-NEXT: v_mov_b32_e32 v10, v1
+; GFX10-NEXT: v_mov_b32_e32 v9, v0
+; GFX10-NEXT: v_add_nc_u32_e32 v15, 0x800, v4
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_readfirstlane_b32 s8, v9
+; GFX10-NEXT: v_readfirstlane_b32 s9, v10
+; GFX10-NEXT: v_readfirstlane_b32 s10, v7
+; GFX10-NEXT: v_readfirstlane_b32 s11, v8
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[9:10]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[7:8]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: buffer_load_dwordx2 v[13:14], v4, s[8:11], 0 offen offset:2048
+; GFX10-NEXT: ; implicit-def: $vgpr4
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB5_1
+; GFX10-NEXT: ; %bb.2:
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: v_max_f64 v[4:5], v[5:6], v[5:6]
+; GFX10-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Loop Header: Depth=1
+; GFX10-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f64 v[0:1], v[13:14], v[13:14]
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_max_f64 v[11:12], v[0:1], v[4:5]
+; GFX10-NEXT: v_mov_b32_e32 v0, v11
+; GFX10-NEXT: v_mov_b32_e32 v1, v12
+; GFX10-NEXT: v_mov_b32_e32 v2, v13
+; GFX10-NEXT: v_mov_b32_e32 v3, v14
+; GFX10-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX10-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX10-NEXT: v_readfirstlane_b32 s8, v9
+; GFX10-NEXT: v_readfirstlane_b32 s9, v10
+; GFX10-NEXT: v_readfirstlane_b32 s10, v7
+; GFX10-NEXT: v_readfirstlane_b32 s11, v8
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[9:10]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[7:8]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v15, s[8:11], 0 offen glc
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB5_4
+; GFX10-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[13:14]
+; GFX10-NEXT: v_mov_b32_e32 v14, v1
+; GFX10-NEXT: v_mov_b32_e32 v13, v0
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB5_3
+; GFX10-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f64__offset__waterfall:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v11, v6
+; GFX90A-NEXT: v_mov_b32_e32 v10, v5
+; GFX90A-NEXT: v_mov_b32_e32 v9, v3
+; GFX90A-NEXT: v_mov_b32_e32 v8, v2
+; GFX90A-NEXT: v_mov_b32_e32 v7, v1
+; GFX90A-NEXT: v_mov_b32_e32 v6, v0
+; GFX90A-NEXT: v_add_u32_e32 v14, 0x800, v4
+; GFX90A-NEXT: s_mov_b64 s[6:7], exec
+; GFX90A-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v6
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v7
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v8
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v9
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[6:7]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[8:9]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_nop 0
+; GFX90A-NEXT: buffer_load_dwordx2 v[12:13], v4, s[8:11], 0 offen offset:2048
+; GFX90A-NEXT: ; implicit-def: $vgpr4
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB5_1
+; GFX90A-NEXT: ; %bb.2:
+; GFX90A-NEXT: s_mov_b64 exec, s[6:7]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_max_f64 v[4:5], v[10:11], v[10:11]
+; GFX90A-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Loop Header: Depth=1
+; GFX90A-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f64 v[0:1], v[12:13], v[12:13]
+; GFX90A-NEXT: v_max_f64 v[10:11], v[0:1], v[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[10:11], v[10:11] op_sel:[0,1]
+; GFX90A-NEXT: s_mov_b64 s[12:13], exec
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[12:13], v[12:13] op_sel:[0,1]
+; GFX90A-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX90A-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v6
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v7
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v8
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v9
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[6:7]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[8:9]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v14, s[8:11], 0 offen glc
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB5_4
+; GFX90A-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX90A-NEXT: s_mov_b64 exec, s[12:13]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[12:13]
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_pk_mov_b32 v[12:13], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB5_3
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f64__offset__waterfall:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v8, v3
+; GFX908-NEXT: v_mov_b32_e32 v7, v2
+; GFX908-NEXT: v_mov_b32_e32 v10, v1
+; GFX908-NEXT: v_mov_b32_e32 v9, v0
+; GFX908-NEXT: v_add_u32_e32 v15, 0x800, v4
+; GFX908-NEXT: s_mov_b64 s[6:7], exec
+; GFX908-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: v_readfirstlane_b32 s8, v9
+; GFX908-NEXT: v_readfirstlane_b32 s9, v10
+; GFX908-NEXT: v_readfirstlane_b32 s10, v7
+; GFX908-NEXT: v_readfirstlane_b32 s11, v8
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_nop 0
+; GFX908-NEXT: buffer_load_dwordx2 v[13:14], v4, s[8:11], 0 offen offset:2048
+; GFX908-NEXT: ; implicit-def: $vgpr4
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_1
+; GFX908-NEXT: ; %bb.2:
+; GFX908-NEXT: s_mov_b64 exec, s[6:7]
+; GFX908-NEXT: v_max_f64 v[4:5], v[5:6], v[5:6]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Loop Header: Depth=1
+; GFX908-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f64 v[0:1], v[13:14], v[13:14]
+; GFX908-NEXT: s_mov_b64 s[12:13], exec
+; GFX908-NEXT: v_max_f64 v[11:12], v[0:1], v[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v11
+; GFX908-NEXT: v_mov_b32_e32 v1, v12
+; GFX908-NEXT: v_mov_b32_e32 v2, v13
+; GFX908-NEXT: v_mov_b32_e32 v3, v14
+; GFX908-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX908-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX908-NEXT: v_readfirstlane_b32 s8, v9
+; GFX908-NEXT: v_readfirstlane_b32 s9, v10
+; GFX908-NEXT: v_readfirstlane_b32 s10, v7
+; GFX908-NEXT: v_readfirstlane_b32 s11, v8
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v15, s[8:11], 0 offen glc
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_4
+; GFX908-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX908-NEXT: s_mov_b64 exec, s[12:13]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[13:14]
+; GFX908-NEXT: v_mov_b32_e32 v14, v1
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v13, v0
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_3
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f64__offset__waterfall:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v8, v3
+; GFX8-NEXT: v_mov_b32_e32 v7, v2
+; GFX8-NEXT: v_mov_b32_e32 v10, v1
+; GFX8-NEXT: v_mov_b32_e32 v9, v0
+; GFX8-NEXT: v_add_u32_e32 v15, vcc, 0x800, v4
+; GFX8-NEXT: s_mov_b64 s[6:7], exec
+; GFX8-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: v_readfirstlane_b32 s8, v9
+; GFX8-NEXT: v_readfirstlane_b32 s9, v10
+; GFX8-NEXT: v_readfirstlane_b32 s10, v7
+; GFX8-NEXT: v_readfirstlane_b32 s11, v8
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_nop 0
+; GFX8-NEXT: buffer_load_dwordx2 v[13:14], v4, s[8:11], 0 offen offset:2048
+; GFX8-NEXT: ; implicit-def: $vgpr4
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_1
+; GFX8-NEXT: ; %bb.2:
+; GFX8-NEXT: s_mov_b64 exec, s[6:7]
+; GFX8-NEXT: v_max_f64 v[4:5], v[5:6], v[5:6]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Loop Header: Depth=1
+; GFX8-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f64 v[0:1], v[13:14], v[13:14]
+; GFX8-NEXT: s_mov_b64 s[12:13], exec
+; GFX8-NEXT: v_max_f64 v[11:12], v[0:1], v[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v11
+; GFX8-NEXT: v_mov_b32_e32 v1, v12
+; GFX8-NEXT: v_mov_b32_e32 v2, v13
+; GFX8-NEXT: v_mov_b32_e32 v3, v14
+; GFX8-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX8-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX8-NEXT: v_readfirstlane_b32 s8, v9
+; GFX8-NEXT: v_readfirstlane_b32 s9, v10
+; GFX8-NEXT: v_readfirstlane_b32 s10, v7
+; GFX8-NEXT: v_readfirstlane_b32 s11, v8
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v15, s[8:11], 0 offen glc
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_4
+; GFX8-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX8-NEXT: s_mov_b64 exec, s[12:13]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[13:14]
+; GFX8-NEXT: v_mov_b32_e32 v14, v1
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v13, v0
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_3
+; GFX8-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f64__offset__waterfall:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v10, v1
+; GFX7-NEXT: v_mov_b32_e32 v9, v0
+; GFX7-NEXT: v_add_i32_e32 v15, vcc, 0x800, v4
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_readfirstlane_b32 s8, v9
+; GFX7-NEXT: v_readfirstlane_b32 s9, v10
+; GFX7-NEXT: v_readfirstlane_b32 s10, v7
+; GFX7-NEXT: v_readfirstlane_b32 s11, v8
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: buffer_load_dwordx2 v[13:14], v4, s[8:11], 0 offen offset:2048
+; GFX7-NEXT: ; implicit-def: $vgpr4
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_1
+; GFX7-NEXT: ; %bb.2:
+; GFX7-NEXT: s_mov_b64 exec, s[6:7]
+; GFX7-NEXT: v_max_f64 v[4:5], v[5:6], v[5:6]
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Loop Header: Depth=1
+; GFX7-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_max_f64 v[0:1], v[13:14], v[13:14]
+; GFX7-NEXT: s_mov_b64 s[12:13], exec
+; GFX7-NEXT: v_max_f64 v[11:12], v[0:1], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v11
+; GFX7-NEXT: v_mov_b32_e32 v1, v12
+; GFX7-NEXT: v_mov_b32_e32 v2, v13
+; GFX7-NEXT: v_mov_b32_e32 v3, v14
+; GFX7-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX7-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX7-NEXT: v_readfirstlane_b32 s8, v9
+; GFX7-NEXT: v_readfirstlane_b32 s9, v10
+; GFX7-NEXT: v_readfirstlane_b32 s10, v7
+; GFX7-NEXT: v_readfirstlane_b32 s11, v8
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v15, s[8:11], 0 offen glc
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_4
+; GFX7-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX7-NEXT: s_mov_b64 exec, s[12:13]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[13:14]
+; GFX7-NEXT: v_mov_b32_e32 v14, v1
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v13, v0
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_3
+; GFX7-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f64__offset__waterfall:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v10, v1
+; GFX6-NEXT: v_mov_b32_e32 v9, v0
+; GFX6-NEXT: v_add_i32_e32 v15, vcc, 0x800, v4
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_readfirstlane_b32 s8, v9
+; GFX6-NEXT: v_readfirstlane_b32 s9, v10
+; GFX6-NEXT: v_readfirstlane_b32 s10, v7
+; GFX6-NEXT: v_readfirstlane_b32 s11, v8
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: buffer_load_dwordx2 v[13:14], v4, s[8:11], 0 offen offset:2048
+; GFX6-NEXT: ; implicit-def: $vgpr4
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB5_1
+; GFX6-NEXT: ; %bb.2:
+; GFX6-NEXT: s_mov_b64 exec, s[6:7]
+; GFX6-NEXT: v_max_f64 v[4:5], v[5:6], v[5:6]
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Loop Header: Depth=1
+; GFX6-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: v_max_f64 v[0:1], v[13:14], v[13:14]
+; GFX6-NEXT: s_mov_b64 s[12:13], exec
+; GFX6-NEXT: v_max_f64 v[11:12], v[0:1], v[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v0, v11
+; GFX6-NEXT: v_mov_b32_e32 v1, v12
+; GFX6-NEXT: v_mov_b32_e32 v2, v13
+; GFX6-NEXT: v_mov_b32_e32 v3, v14
+; GFX6-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX6-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX6-NEXT: v_readfirstlane_b32 s8, v9
+; GFX6-NEXT: v_readfirstlane_b32 s9, v10
+; GFX6-NEXT: v_readfirstlane_b32 s10, v7
+; GFX6-NEXT: v_readfirstlane_b32 s11, v8
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v15, s[8:11], 0 offen glc
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB5_4
+; GFX6-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX6-NEXT: s_mov_b64 exec, s[12:13]
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[13:14]
+; GFX6-NEXT: v_mov_b32_e32 v14, v1
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v13, v0
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB5_3
+; GFX6-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmax ptr addrspace(7) %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+; --------------------------------------------------------------------
+; half
+; --------------------------------------------------------------------
+
+define half @buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset(ptr addrspace(7) inreg %ptr, half %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_addk_co_i32 s4, 0x200
+; GFX12-NEXT: v_max_num_f16_e32 v5, v0, v0
+; GFX12-NEXT: s_and_b32 s4, s4, -4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_mov_b32_e32 v4, s4
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: buffer_load_b32 v1, v4, s[0:3], null offen
+; GFX12-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v0, v0, v0
+; GFX12-NEXT: v_max_num_f16_e32 v0, v0, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX12-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[2:3], v4, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB6_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_addk_i32 s4, 0x200
+; GFX940-NEXT: s_and_b32 s4, s4, -4
+; GFX940-NEXT: v_mov_b32_e32 v4, s4
+; GFX940-NEXT: buffer_load_dword v1, v4, s[0:3], 0 offen
+; GFX940-NEXT: s_mov_b32 s6, 0xffffff
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_max_f16_e32 v5, v0, v0
+; GFX940-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX940-NEXT: v_max_f16_sdwa v0, v0, v5 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_and_or_b32 v0, v1, s6, v0
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], v[0:1]
+; GFX940-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB6_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_addk_i32 s4, 0x200
+; GFX11-NEXT: v_max_f16_e32 v5, v0, v0
+; GFX11-NEXT: s_and_b32 s4, s4, -4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v4, s4
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: buffer_load_b32 v1, v4, s[0:3], 0 offen
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v0, v0, v0
+; GFX11-NEXT: v_max_f16_e32 v0, v0, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX11-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[2:3], v4, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB6_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_addk_i32 s8, 0x200
+; GFX10-NEXT: v_max_f16_e32 v5, v0, v0
+; GFX10-NEXT: s_and_b32 s8, s8, -4
+; GFX10-NEXT: v_mov_b32_e32 v4, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX10-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_max_f16_e32 v0, v0, v5
+; GFX10-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX10-NEXT: v_mov_b32_e32 v3, v1
+; GFX10-NEXT: v_mov_b32_e32 v2, v0
+; GFX10-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB6_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: s_addk_i32 s8, 0x200
+; GFX90A-NEXT: s_and_b32 s8, s8, -4
+; GFX90A-NEXT: v_mov_b32_e32 v4, s8
+; GFX90A-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX90A-NEXT: s_mov_b32 s10, 0xffffff
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_max_f16_e32 v5, v0, v0
+; GFX90A-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX90A-NEXT: v_max_f16_sdwa v0, v0, v5 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_and_or_b32 v0, v1, s10, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB6_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: s_addk_i32 s8, 0x200
+; GFX908-NEXT: s_and_b32 s8, s8, -4
+; GFX908-NEXT: v_mov_b32_e32 v4, s8
+; GFX908-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX908-NEXT: s_mov_b32 s10, 0xffffff
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_max_f16_e32 v5, v0, v0
+; GFX908-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX908-NEXT: v_max_f16_sdwa v0, v0, v5 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_and_or_b32 v0, v1, s10, v0
+; GFX908-NEXT: v_mov_b32_e32 v3, v1
+; GFX908-NEXT: v_mov_b32_e32 v2, v0
+; GFX908-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB6_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_addk_i32 s8, 0x200
+; GFX8-NEXT: s_and_b32 s8, s8, -4
+; GFX8-NEXT: v_mov_b32_e32 v4, s8
+; GFX8-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_max_f16_e32 v5, v0, v0
+; GFX8-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffffff, v1
+; GFX8-NEXT: v_max_f16_sdwa v0, v0, v5 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_or_b32_e32 v0, v2, v0
+; GFX8-NEXT: v_mov_b32_e32 v3, v1
+; GFX8-NEXT: v_mov_b32_e32 v2, v0
+; GFX8-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB6_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_addk_i32 s8, 0x200
+; GFX7-NEXT: s_and_b32 s8, s8, -4
+; GFX7-NEXT: v_mov_b32_e32 v4, s8
+; GFX7-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v0
+; GFX7-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffffff, v1
+; GFX7-NEXT: v_max_f32_e32 v0, v0, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX7-NEXT: v_or_b32_e32 v0, v2, v0
+; GFX7-NEXT: v_mov_b32_e32 v3, v1
+; GFX7-NEXT: v_mov_b32_e32 v2, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB6_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_addk_i32 s8, 0x200
+; GFX6-NEXT: s_and_b32 s8, s8, -4
+; GFX6-NEXT: v_mov_b32_e32 v4, s8
+; GFX6-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v0
+; GFX6-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffffff, v1
+; GFX6-NEXT: v_max_f32_e32 v0, v0, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX6-NEXT: v_or_b32_e32 v0, v2, v0
+; GFX6-NEXT: v_mov_b32_e32 v3, v1
+; GFX6-NEXT: v_mov_b32_e32 v2, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB6_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmax ptr addrspace(7) %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+define void @buffer_fat_ptr_agent_atomic_fmax_noret_f16__offset(ptr addrspace(7) inreg %ptr, half %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_addk_co_i32 s4, 0x200
+; GFX12-NEXT: v_max_num_f16_e32 v3, v0, v0
+; GFX12-NEXT: s_and_b32 s4, s4, -4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_mov_b32_e32 v2, s4
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: buffer_load_b32 v1, v2, s[0:3], null offen
+; GFX12-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v0, v0, v0
+; GFX12-NEXT: v_max_num_f16_e32 v0, v0, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX12-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v2, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v4
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB7_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_addk_i32 s4, 0x200
+; GFX940-NEXT: s_and_b32 s4, s4, -4
+; GFX940-NEXT: v_mov_b32_e32 v2, s4
+; GFX940-NEXT: buffer_load_dword v1, v2, s[0:3], 0 offen
+; GFX940-NEXT: s_mov_b32 s6, 0xffffff
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_max_f16_e32 v3, v0, v0
+; GFX940-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX940-NEXT: v_max_f16_sdwa v0, v0, v3 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_and_or_b32 v0, v1, s6, v0
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[0:1]
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB7_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_addk_i32 s4, 0x200
+; GFX11-NEXT: v_max_f16_e32 v3, v0, v0
+; GFX11-NEXT: s_and_b32 s4, s4, -4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v2, s4
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: buffer_load_b32 v1, v2, s[0:3], 0 offen
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v0, v0, v0
+; GFX11-NEXT: v_max_f16_e32 v0, v0, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX11-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v2, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v4
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB7_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_addk_i32 s8, 0x200
+; GFX10-NEXT: v_max_f16_e32 v3, v0, v0
+; GFX10-NEXT: s_and_b32 s8, s8, -4
+; GFX10-NEXT: v_mov_b32_e32 v2, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX10-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_max_f16_e32 v0, v0, v3
+; GFX10-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX10-NEXT: v_mov_b32_e32 v5, v1
+; GFX10-NEXT: v_mov_b32_e32 v4, v0
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v4
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB7_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: s_addk_i32 s8, 0x200
+; GFX90A-NEXT: s_and_b32 s8, s8, -4
+; GFX90A-NEXT: v_mov_b32_e32 v2, s8
+; GFX90A-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX90A-NEXT: s_mov_b32 s10, 0xffffff
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_max_f16_e32 v3, v0, v0
+; GFX90A-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX90A-NEXT: v_max_f16_sdwa v0, v0, v3 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_and_or_b32 v0, v1, s10, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB7_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: s_addk_i32 s8, 0x200
+; GFX908-NEXT: s_and_b32 s8, s8, -4
+; GFX908-NEXT: v_mov_b32_e32 v2, s8
+; GFX908-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX908-NEXT: s_mov_b32 s10, 0xffffff
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_max_f16_e32 v3, v0, v0
+; GFX908-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX908-NEXT: v_max_f16_sdwa v0, v0, v3 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_and_or_b32 v0, v1, s10, v0
+; GFX908-NEXT: v_mov_b32_e32 v5, v1
+; GFX908-NEXT: v_mov_b32_e32 v4, v0
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: v_mov_b32_e32 v1, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB7_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_addk_i32 s8, 0x200
+; GFX8-NEXT: s_and_b32 s8, s8, -4
+; GFX8-NEXT: v_mov_b32_e32 v2, s8
+; GFX8-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_max_f16_e32 v3, v0, v0
+; GFX8-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX8-NEXT: v_and_b32_e32 v4, 0xffffff, v1
+; GFX8-NEXT: v_max_f16_sdwa v0, v0, v3 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX8-NEXT: v_mov_b32_e32 v5, v1
+; GFX8-NEXT: v_mov_b32_e32 v4, v0
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v1, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB7_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_addk_i32 s8, 0x200
+; GFX7-NEXT: s_and_b32 s8, s8, -4
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v0
+; GFX7-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffffff, v1
+; GFX7-NEXT: v_max_f32_e32 v0, v0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX7-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX7-NEXT: v_mov_b32_e32 v5, v1
+; GFX7-NEXT: v_mov_b32_e32 v4, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v1, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB7_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_addk_i32 s8, 0x200
+; GFX6-NEXT: s_and_b32 s8, s8, -4
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v0
+; GFX6-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffffff, v1
+; GFX6-NEXT: v_max_f32_e32 v0, v0, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX6-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX6-NEXT: v_mov_b32_e32 v5, v1
+; GFX6-NEXT: v_mov_b32_e32 v4, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v1, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB7_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(7) %ptr, i32 256
+ %unused = atomicrmw fmax ptr addrspace(7) %gep, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define half @buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset__waterfall(ptr addrspace(7) %ptr, half %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset__waterfall:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v4, 0x200, v4
+; GFX12-NEXT: s_mov_b32 s1, exec_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX12-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: buffer_load_b32 v6, v7, s[4:7], null offen
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB8_1
+; GFX12-NEXT: ; %bb.2:
+; GFX12-NEXT: s_mov_b32 exec_lo, s1
+; GFX12-NEXT: v_max_num_f16_e32 v8, v5, v5
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Loop Header: Depth=1
+; GFX12-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX12-NEXT: s_mov_b32 s2, exec_lo
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v4, v4, v4
+; GFX12-NEXT: v_max_num_f16_e32 v4, v4, v8
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, 0xffffff, v6, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v5
+; GFX12-NEXT: v_mov_b32_e32 v5, v6
+; GFX12-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX12-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB8_4
+; GFX12-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX12-NEXT: s_mov_b32 exec_lo, s2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX12-NEXT: v_mov_b32_e32 v6, v4
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB8_3
+; GFX12-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset__waterfall:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v4, 0x200, v4
+; GFX940-NEXT: v_and_b32_e32 v8, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v9, 0xffffff
+; GFX940-NEXT: s_mov_b64 s[2:3], exec
+; GFX940-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: buffer_load_dword v7, v8, s[4:7], 0 offen
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB8_1
+; GFX940-NEXT: ; %bb.2:
+; GFX940-NEXT: s_mov_b64 exec, s[2:3]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_max_f16_e32 v10, v5, v5
+; GFX940-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Loop Header: Depth=1
+; GFX940-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f16_sdwa v4, v7, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX940-NEXT: v_max_f16_sdwa v4, v4, v10 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_mov_b64 s[8:9], exec
+; GFX940-NEXT: v_and_or_b32 v6, v7, v9, v4
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[6:7]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX940-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[4:7], 0 offen sc0
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB8_4
+; GFX940-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX940-NEXT: s_mov_b64 exec, s[8:9]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB8_3
+; GFX940-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset__waterfall:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v4, 0x200, v4
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX11-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: buffer_load_b32 v6, v7, s[4:7], 0 offen
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB8_1
+; GFX11-NEXT: ; %bb.2:
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: v_max_f16_e32 v8, v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Loop Header: Depth=1
+; GFX11-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX11-NEXT: v_max_f16_e32 v4, v4, v8
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, 0xffffff, v6, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v5
+; GFX11-NEXT: v_mov_b32_e32 v5, v6
+; GFX11-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX11-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], 0 offen glc
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB8_4
+; GFX11-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX11-NEXT: v_mov_b32_e32 v6, v4
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB8_3
+; GFX11-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset__waterfall:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v4, 0x200, v4
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX10-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB8_1
+; GFX10-NEXT: ; %bb.2:
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: v_max_f16_e32 v8, v5, v5
+; GFX10-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Loop Header: Depth=1
+; GFX10-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f16_sdwa v4, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_max_f16_e32 v4, v4, v8
+; GFX10-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX10-NEXT: v_and_or_b32 v5, 0xffffff, v6, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v5
+; GFX10-NEXT: v_mov_b32_e32 v5, v6
+; GFX10-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX10-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB8_4
+; GFX10-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB8_3
+; GFX10-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset__waterfall:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v4, 0x200, v4
+; GFX90A-NEXT: v_and_b32_e32 v8, -4, v4
+; GFX90A-NEXT: v_mov_b32_e32 v9, 0xffffff
+; GFX90A-NEXT: s_mov_b64 s[6:7], exec
+; GFX90A-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_nop 0
+; GFX90A-NEXT: buffer_load_dword v7, v8, s[8:11], 0 offen
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB8_1
+; GFX90A-NEXT: ; %bb.2:
+; GFX90A-NEXT: s_mov_b64 exec, s[6:7]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_max_f16_e32 v10, v5, v5
+; GFX90A-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Loop Header: Depth=1
+; GFX90A-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f16_sdwa v4, v7, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX90A-NEXT: v_max_f16_sdwa v4, v4, v10 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v9, v4
+; GFX90A-NEXT: s_mov_b64 s[12:13], exec
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[6:7] op_sel:[0,1]
+; GFX90A-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX90A-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[8:11], 0 offen glc
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB8_4
+; GFX90A-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX90A-NEXT: s_mov_b64 exec, s[12:13]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB8_3
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset__waterfall:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v4, 0x200, v4
+; GFX908-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX908-NEXT: v_mov_b32_e32 v8, 0xffffff
+; GFX908-NEXT: s_mov_b64 s[6:7], exec
+; GFX908-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_nop 0
+; GFX908-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_1
+; GFX908-NEXT: ; %bb.2:
+; GFX908-NEXT: s_mov_b64 exec, s[6:7]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_max_f16_e32 v9, v5, v5
+; GFX908-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Loop Header: Depth=1
+; GFX908-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f16_sdwa v4, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX908-NEXT: v_max_f16_sdwa v4, v4, v9 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_and_or_b32 v5, v6, v8, v4
+; GFX908-NEXT: v_mov_b32_e32 v4, v5
+; GFX908-NEXT: s_mov_b64 s[12:13], exec
+; GFX908-NEXT: v_mov_b32_e32 v5, v6
+; GFX908-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX908-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_4
+; GFX908-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX908-NEXT: s_mov_b64 exec, s[12:13]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_3
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset__waterfall:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x200, v4
+; GFX8-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX8-NEXT: s_mov_b64 s[6:7], exec
+; GFX8-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_nop 0
+; GFX8-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_1
+; GFX8-NEXT: ; %bb.2:
+; GFX8-NEXT: s_mov_b64 exec, s[6:7]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_max_f16_e32 v8, v5, v5
+; GFX8-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Loop Header: Depth=1
+; GFX8-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v4, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX8-NEXT: v_max_f16_sdwa v4, v4, v8 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffffff, v6
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v4
+; GFX8-NEXT: v_mov_b32_e32 v4, v5
+; GFX8-NEXT: s_mov_b64 s[12:13], exec
+; GFX8-NEXT: v_mov_b32_e32 v5, v6
+; GFX8-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX8-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_4
+; GFX8-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX8-NEXT: s_mov_b64 exec, s[12:13]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_3
+; GFX8-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset__waterfall:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x200, v4
+; GFX7-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB8_1
+; GFX7-NEXT: ; %bb.2:
+; GFX7-NEXT: s_mov_b64 exec, s[6:7]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v5
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v8, v4
+; GFX7-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Loop Header: Depth=1
+; GFX7-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffffff, v6
+; GFX7-NEXT: s_mov_b64 s[12:13], exec
+; GFX7-NEXT: v_max_f32_e32 v4, v4, v8
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX7-NEXT: v_or_b32_e32 v5, v5, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: v_mov_b32_e32 v5, v6
+; GFX7-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX7-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB8_4
+; GFX7-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX7-NEXT: s_mov_b64 exec, s[12:13]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB8_3
+; GFX7-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset__waterfall:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x200, v4
+; GFX6-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB8_1
+; GFX6-NEXT: ; %bb.2:
+; GFX6-NEXT: s_mov_b64 exec, s[6:7]
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v5
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v8, v4
+; GFX6-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Loop Header: Depth=1
+; GFX6-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffffff, v6
+; GFX6-NEXT: s_mov_b64 s[12:13], exec
+; GFX6-NEXT: v_max_f32_e32 v4, v4, v8
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX6-NEXT: v_or_b32_e32 v5, v5, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: v_mov_b32_e32 v5, v6
+; GFX6-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX6-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB8_4
+; GFX6-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX6-NEXT: s_mov_b64 exec, s[12:13]
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB8_3
+; GFX6-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmax ptr addrspace(7) %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+; --------------------------------------------------------------------
+; bfloat
+; --------------------------------------------------------------------
+
+define bfloat @buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset(ptr addrspace(7) inreg %ptr, bfloat %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_addk_co_i32 s4, 0x200
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX12-NEXT: s_and_b32 s4, s4, -4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_mov_b32_e32 v4, s4
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: buffer_load_b32 v1, v4, s[0:3], null offen
+; GFX12-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX12-NEXT: v_max_num_f32_e32 v0, v0, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v3, 0x400000, v0
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX12-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX12-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[2:3], v4, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB9_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_addk_i32 s4, 0x200
+; GFX940-NEXT: s_and_b32 s4, s4, -4
+; GFX940-NEXT: v_mov_b32_e32 v4, s4
+; GFX940-NEXT: buffer_load_dword v1, v4, s[0:3], 0 offen
+; GFX940-NEXT: s_mov_b32 s6, 0xffffff
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX940-NEXT: s_movk_i32 s7, 0x7fff
+; GFX940-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX940-NEXT: v_max_f32_e32 v0, v0, v5
+; GFX940-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v3, 0x400000, v0
+; GFX940-NEXT: v_add3_u32 v2, v2, v0, s7
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX940-NEXT: v_and_or_b32 v0, v1, s6, v0
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], v[0:1]
+; GFX940-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB9_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_addk_i32 s4, 0x200
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX11-NEXT: s_and_b32 s4, s4, -4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v4, s4
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: buffer_load_b32 v1, v4, s[0:3], 0 offen
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT: v_max_f32_e32 v0, v0, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v3, 0x400000, v0
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX11-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[2:3], v4, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB9_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_addk_i32 s8, 0x200
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX10-NEXT: s_and_b32 s8, s8, -4
+; GFX10-NEXT: v_mov_b32_e32 v4, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX10-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX10-NEXT: v_max_f32_e32 v0, v0, v5
+; GFX10-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v3, 0x400000, v0
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX10-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX10-NEXT: v_mov_b32_e32 v3, v1
+; GFX10-NEXT: v_mov_b32_e32 v2, v0
+; GFX10-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB9_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: s_addk_i32 s8, 0x200
+; GFX90A-NEXT: s_and_b32 s8, s8, -4
+; GFX90A-NEXT: v_mov_b32_e32 v4, s8
+; GFX90A-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX90A-NEXT: s_mov_b32 s10, 0xffffff
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX90A-NEXT: s_movk_i32 s11, 0x7fff
+; GFX90A-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX90A-NEXT: v_max_f32_e32 v0, v0, v5
+; GFX90A-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v3, 0x400000, v0
+; GFX90A-NEXT: v_add3_u32 v2, v2, v0, s11
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX90A-NEXT: v_and_or_b32 v0, v1, s10, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB9_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: s_addk_i32 s8, 0x200
+; GFX908-NEXT: s_and_b32 s8, s8, -4
+; GFX908-NEXT: v_mov_b32_e32 v4, s8
+; GFX908-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX908-NEXT: s_mov_b32 s10, 0xffffff
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX908-NEXT: s_movk_i32 s11, 0x7fff
+; GFX908-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX908-NEXT: v_max_f32_e32 v0, v0, v5
+; GFX908-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v3, 0x400000, v0
+; GFX908-NEXT: v_add3_u32 v2, v2, v0, s11
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX908-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX908-NEXT: v_and_or_b32 v0, v1, s10, v0
+; GFX908-NEXT: v_mov_b32_e32 v3, v1
+; GFX908-NEXT: v_mov_b32_e32 v2, v0
+; GFX908-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB9_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_addk_i32 s8, 0x200
+; GFX8-NEXT: s_and_b32 s8, s8, -4
+; GFX8-NEXT: v_mov_b32_e32 v4, s8
+; GFX8-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX8-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: v_max_f32_e32 v0, v0, v5
+; GFX8-NEXT: v_bfe_u32 v3, v0, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v0
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
+; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v0, v3, v6, vcc
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffffff, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX8-NEXT: v_or_b32_e32 v0, v2, v0
+; GFX8-NEXT: v_mov_b32_e32 v3, v1
+; GFX8-NEXT: v_mov_b32_e32 v2, v0
+; GFX8-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB9_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_addk_i32 s8, 0x200
+; GFX7-NEXT: s_and_b32 s8, s8, -4
+; GFX7-NEXT: v_mov_b32_e32 v4, s8
+; GFX7-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v0
+; GFX7-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_max_f32_e32 v0, v0, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffffff, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX7-NEXT: v_or_b32_e32 v0, v2, v0
+; GFX7-NEXT: v_mov_b32_e32 v3, v1
+; GFX7-NEXT: v_mov_b32_e32 v2, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB9_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_addk_i32 s8, 0x200
+; GFX6-NEXT: s_and_b32 s8, s8, -4
+; GFX6-NEXT: v_mov_b32_e32 v4, s8
+; GFX6-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v0
+; GFX6-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: v_max_f32_e32 v0, v0, v5
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffffff, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX6-NEXT: v_or_b32_e32 v0, v2, v0
+; GFX6-NEXT: v_mov_b32_e32 v3, v1
+; GFX6-NEXT: v_mov_b32_e32 v2, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB9_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmax ptr addrspace(7) %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+define void @buffer_fat_ptr_agent_atomic_fmax_noret_bf16__offset(ptr addrspace(7) inreg %ptr, bfloat %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_addk_co_i32 s4, 0x200
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX12-NEXT: s_and_b32 s4, s4, -4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_mov_b32_e32 v2, s4
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: buffer_load_b32 v1, v2, s[0:3], null offen
+; GFX12-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX12-NEXT: v_max_num_f32_e32 v0, v0, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v4, v0, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX12-NEXT: v_add3_u32 v4, v4, v0, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX12-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v2, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v4
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB10_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_addk_i32 s4, 0x200
+; GFX940-NEXT: s_and_b32 s4, s4, -4
+; GFX940-NEXT: v_mov_b32_e32 v2, s4
+; GFX940-NEXT: buffer_load_dword v1, v2, s[0:3], 0 offen
+; GFX940-NEXT: s_mov_b32 s6, 0xffffff
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX940-NEXT: s_movk_i32 s7, 0x7fff
+; GFX940-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX940-NEXT: v_max_f32_e32 v0, v0, v3
+; GFX940-NEXT: v_bfe_u32 v4, v0, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX940-NEXT: v_add3_u32 v4, v4, v0, s7
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX940-NEXT: v_and_or_b32 v0, v1, s6, v0
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[0:1]
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB10_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_addk_i32 s4, 0x200
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX11-NEXT: s_and_b32 s4, s4, -4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v2, s4
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: buffer_load_b32 v1, v2, s[0:3], 0 offen
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT: v_max_f32_e32 v0, v0, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v4, v0, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT: v_add3_u32 v4, v4, v0, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX11-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v2, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v4
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB10_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_addk_i32 s8, 0x200
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX10-NEXT: s_and_b32 s8, s8, -4
+; GFX10-NEXT: v_mov_b32_e32 v2, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX10-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX10-NEXT: v_max_f32_e32 v0, v0, v3
+; GFX10-NEXT: v_bfe_u32 v4, v0, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX10-NEXT: v_add3_u32 v4, v4, v0, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX10-NEXT: v_mov_b32_e32 v5, v1
+; GFX10-NEXT: v_mov_b32_e32 v4, v0
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v4
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB10_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: s_addk_i32 s8, 0x200
+; GFX90A-NEXT: s_and_b32 s8, s8, -4
+; GFX90A-NEXT: v_mov_b32_e32 v2, s8
+; GFX90A-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX90A-NEXT: s_mov_b32 s10, 0xffffff
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX90A-NEXT: s_movk_i32 s11, 0x7fff
+; GFX90A-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX90A-NEXT: v_max_f32_e32 v0, v0, v3
+; GFX90A-NEXT: v_bfe_u32 v4, v0, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX90A-NEXT: v_add3_u32 v4, v4, v0, s11
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX90A-NEXT: v_and_or_b32 v0, v1, s10, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB10_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: s_addk_i32 s8, 0x200
+; GFX908-NEXT: s_and_b32 s8, s8, -4
+; GFX908-NEXT: v_mov_b32_e32 v2, s8
+; GFX908-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX908-NEXT: s_mov_b32 s10, 0xffffff
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX908-NEXT: s_movk_i32 s11, 0x7fff
+; GFX908-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX908-NEXT: v_max_f32_e32 v0, v0, v3
+; GFX908-NEXT: v_bfe_u32 v4, v0, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX908-NEXT: v_add3_u32 v4, v4, v0, s11
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX908-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX908-NEXT: v_and_or_b32 v0, v1, s10, v0
+; GFX908-NEXT: v_mov_b32_e32 v5, v1
+; GFX908-NEXT: v_mov_b32_e32 v4, v0
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: v_mov_b32_e32 v1, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB10_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_addk_i32 s8, 0x200
+; GFX8-NEXT: s_and_b32 s8, s8, -4
+; GFX8-NEXT: v_mov_b32_e32 v2, s8
+; GFX8-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX8-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: v_max_f32_e32 v0, v0, v3
+; GFX8-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v0
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
+; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v0, v5, v6, vcc
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: v_and_b32_e32 v4, 0xffffff, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX8-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX8-NEXT: v_mov_b32_e32 v5, v1
+; GFX8-NEXT: v_mov_b32_e32 v4, v0
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v1, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB10_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_addk_i32 s8, 0x200
+; GFX7-NEXT: s_and_b32 s8, s8, -4
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX7-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_max_f32_e32 v0, v0, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffffff, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX7-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX7-NEXT: v_mov_b32_e32 v5, v1
+; GFX7-NEXT: v_mov_b32_e32 v4, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v1, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB10_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_addk_i32 s8, 0x200
+; GFX6-NEXT: s_and_b32 s8, s8, -4
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX6-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: v_max_f32_e32 v0, v0, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffffff, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX6-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX6-NEXT: v_mov_b32_e32 v5, v1
+; GFX6-NEXT: v_mov_b32_e32 v4, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v1, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB10_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(7) %ptr, i32 256
+ %unused = atomicrmw fmax ptr addrspace(7) %gep, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define bfloat @buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__waterfall(ptr addrspace(7) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__waterfall:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v4, 0x200, v4
+; GFX12-NEXT: s_mov_b32 s1, exec_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX12-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: buffer_load_b32 v6, v7, s[4:7], null offen
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB11_1
+; GFX12-NEXT: ; %bb.2:
+; GFX12-NEXT: s_mov_b32 exec_lo, s1
+; GFX12-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Loop Header: Depth=1
+; GFX12-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX12-NEXT: s_mov_b32 s2, exec_lo
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: v_max_num_f32_e32 v4, v4, v8
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v4
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX12-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v4, v5, v9, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX12-NEXT: v_and_or_b32 v5, 0xffffff, v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_mov_b32_e32 v4, v5
+; GFX12-NEXT: v_mov_b32_e32 v5, v6
+; GFX12-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX12-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB11_4
+; GFX12-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX12-NEXT: s_mov_b32 exec_lo, s2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX12-NEXT: v_mov_b32_e32 v6, v4
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB11_3
+; GFX12-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__waterfall:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v4, 0x200, v4
+; GFX940-NEXT: v_and_b32_e32 v8, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v9, 0xffffff
+; GFX940-NEXT: s_mov_b64 s[2:3], exec
+; GFX940-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: buffer_load_dword v7, v8, s[4:7], 0 offen
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB11_1
+; GFX940-NEXT: ; %bb.2:
+; GFX940-NEXT: s_mov_b64 exec, s[2:3]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v10, 16, v5
+; GFX940-NEXT: s_movk_i32 s10, 0x7fff
+; GFX940-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Loop Header: Depth=1
+; GFX940-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, 24, v7
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX940-NEXT: v_max_f32_e32 v4, v4, v10
+; GFX940-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX940-NEXT: v_add3_u32 v5, v5, v4, s10
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_mov_b64 s[8:9], exec
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX940-NEXT: v_and_or_b32 v6, v7, v9, v4
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[6:7]
+; GFX940-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX940-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[4:7], 0 offen sc0
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB11_4
+; GFX940-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX940-NEXT: s_mov_b64 exec, s[8:9]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB11_3
+; GFX940-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__waterfall:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v4, 0x200, v4
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX11-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: buffer_load_b32 v6, v7, s[4:7], 0 offen
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB11_1
+; GFX11-NEXT: ; %bb.2:
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Loop Header: Depth=1
+; GFX11-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: v_max_f32_e32 v4, v4, v8
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v4
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v4, v5, v9, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX11-NEXT: v_and_or_b32 v5, 0xffffff, v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_mov_b32_e32 v4, v5
+; GFX11-NEXT: v_mov_b32_e32 v5, v6
+; GFX11-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX11-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], 0 offen glc
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB11_4
+; GFX11-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX11-NEXT: v_mov_b32_e32 v6, v4
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB11_3
+; GFX11-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__waterfall:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v4, 0x200, v4
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX10-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB11_1
+; GFX10-NEXT: ; %bb.2:
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX10-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Loop Header: Depth=1
+; GFX10-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX10-NEXT: v_max_f32_e32 v4, v4, v8
+; GFX10-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v4
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX10-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v4, v5, v9, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX10-NEXT: v_and_or_b32 v5, 0xffffff, v6, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v5
+; GFX10-NEXT: v_mov_b32_e32 v5, v6
+; GFX10-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX10-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB11_4
+; GFX10-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB11_3
+; GFX10-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__waterfall:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v4, 0x200, v4
+; GFX90A-NEXT: v_and_b32_e32 v8, -4, v4
+; GFX90A-NEXT: v_mov_b32_e32 v9, 0xffffff
+; GFX90A-NEXT: s_mov_b64 s[6:7], exec
+; GFX90A-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_nop 0
+; GFX90A-NEXT: buffer_load_dword v7, v8, s[8:11], 0 offen
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_1
+; GFX90A-NEXT: ; %bb.2:
+; GFX90A-NEXT: s_mov_b64 exec, s[6:7]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v10, 16, v5
+; GFX90A-NEXT: s_movk_i32 s14, 0x7fff
+; GFX90A-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Loop Header: Depth=1
+; GFX90A-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, 24, v7
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX90A-NEXT: v_max_f32_e32 v4, v4, v10
+; GFX90A-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX90A-NEXT: v_add3_u32 v5, v5, v4, s14
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v9, v4
+; GFX90A-NEXT: s_mov_b64 s[12:13], exec
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[6:7] op_sel:[0,1]
+; GFX90A-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX90A-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[8:11], 0 offen glc
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_4
+; GFX90A-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX90A-NEXT: s_mov_b64 exec, s[12:13]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_3
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__waterfall:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v4, 0x200, v4
+; GFX908-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX908-NEXT: v_mov_b32_e32 v8, 0xffffff
+; GFX908-NEXT: s_mov_b64 s[6:7], exec
+; GFX908-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_nop 0
+; GFX908-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_1
+; GFX908-NEXT: ; %bb.2:
+; GFX908-NEXT: s_mov_b64 exec, s[6:7]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v9, 16, v5
+; GFX908-NEXT: s_movk_i32 s14, 0x7fff
+; GFX908-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Loop Header: Depth=1
+; GFX908-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX908-NEXT: v_max_f32_e32 v4, v4, v9
+; GFX908-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX908-NEXT: v_add3_u32 v5, v5, v4, s14
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v4
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX908-NEXT: v_cndmask_b32_e32 v4, v5, v10, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX908-NEXT: v_and_or_b32 v5, v6, v8, v4
+; GFX908-NEXT: v_mov_b32_e32 v4, v5
+; GFX908-NEXT: s_mov_b64 s[12:13], exec
+; GFX908-NEXT: v_mov_b32_e32 v5, v6
+; GFX908-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX908-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_4
+; GFX908-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX908-NEXT: s_mov_b64 exec, s[12:13]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_3
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__waterfall:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x200, v4
+; GFX8-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX8-NEXT: s_mov_b64 s[6:7], exec
+; GFX8-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_nop 0
+; GFX8-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_1
+; GFX8-NEXT: ; %bb.2:
+; GFX8-NEXT: s_mov_b64 exec, s[6:7]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX8-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Loop Header: Depth=1
+; GFX8-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX8-NEXT: v_max_f32_e32 v4, v4, v8
+; GFX8-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v4
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v4, v5, v9, vcc
+; GFX8-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffffff, v6
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v4
+; GFX8-NEXT: v_mov_b32_e32 v4, v5
+; GFX8-NEXT: s_mov_b64 s[12:13], exec
+; GFX8-NEXT: v_mov_b32_e32 v5, v6
+; GFX8-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX8-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_4
+; GFX8-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX8-NEXT: s_mov_b64 exec, s[12:13]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_3
+; GFX8-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__waterfall:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x200, v4
+; GFX7-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB11_1
+; GFX7-NEXT: ; %bb.2:
+; GFX7-NEXT: s_mov_b64 exec, s[6:7]
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v5
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v4
+; GFX7-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Loop Header: Depth=1
+; GFX7-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_max_f32_e32 v4, v4, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffffff, v6
+; GFX7-NEXT: v_or_b32_e32 v5, v5, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_mov_b64 s[12:13], exec
+; GFX7-NEXT: v_mov_b32_e32 v5, v6
+; GFX7-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX7-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB11_4
+; GFX7-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX7-NEXT: s_mov_b64 exec, s[12:13]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB11_3
+; GFX7-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__waterfall:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x200, v4
+; GFX6-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB11_1
+; GFX6-NEXT: ; %bb.2:
+; GFX6-NEXT: s_mov_b64 exec, s[6:7]
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v5
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: v_and_b32_e32 v8, 0xffff0000, v4
+; GFX6-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Loop Header: Depth=1
+; GFX6-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_max_f32_e32 v4, v4, v8
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffffff, v6
+; GFX6-NEXT: v_or_b32_e32 v5, v5, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_mov_b64 s[12:13], exec
+; GFX6-NEXT: v_mov_b32_e32 v5, v6
+; GFX6-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX6-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB11_4
+; GFX6-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX6-NEXT: s_mov_b64 exec, s[12:13]
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB11_3
+; GFX6-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmax ptr addrspace(7) %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+; --------------------------------------------------------------------
+; <2 x half>
+; --------------------------------------------------------------------
+
+define <2 x half> @buffer_fat_ptr_agent_atomic_fmax_ret_v2f16__offset(ptr addrspace(7) inreg %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s4
+; GFX12-NEXT: s_addk_co_i32 s4, 0x400
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_mov_b32_e32 v3, s4
+; GFX12-NEXT: v_pk_max_num_f16 v2, v1, v1
+; GFX12-NEXT: buffer_load_b32 v0, v0, s[0:3], null offen offset:1024
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v5, v0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v0, v5, v5
+; GFX12-NEXT: v_pk_max_num_f16 v4, v0, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v3, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB12_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, v0
+; GFX940-NEXT: v_mov_b32_e32 v0, s4
+; GFX940-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen offset:1024
+; GFX940-NEXT: s_add_i32 s6, s4, 0x400
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_pk_max_f16 v2, v1, v1
+; GFX940-NEXT: v_mov_b32_e32 v3, s6
+; GFX940-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v0
+; GFX940-NEXT: v_pk_max_f16 v0, v5, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_pk_max_f16 v4, v0, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_mov_b64_e32 v[0:1], v[4:5]
+; GFX940-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB12_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s4
+; GFX11-NEXT: s_addk_i32 s4, 0x400
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_mov_b32_e32 v3, s4
+; GFX11-NEXT: v_pk_max_f16 v2, v1, v1
+; GFX11-NEXT: buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v5, v0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v0, v5, v5
+; GFX11-NEXT: v_pk_max_f16 v4, v0, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v3, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB12_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_mov_b32_e32 v0, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: v_mov_b32_e32 v3, s8
+; GFX10-NEXT: v_pk_max_f16 v2, v1, v1
+; GFX10-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v5, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_pk_max_f16 v0, v5, v5
+; GFX10-NEXT: v_pk_max_f16 v4, v0, v2
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB12_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: v_mov_b32_e32 v0, s8
+; GFX90A-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX90A-NEXT: s_add_i32 s10, s8, 0x400
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_pk_max_f16 v2, v1, v1
+; GFX90A-NEXT: v_mov_b32_e32 v3, s10
+; GFX90A-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v0
+; GFX90A-NEXT: v_pk_max_f16 v0, v5, v5
+; GFX90A-NEXT: v_pk_max_f16 v4, v0, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB12_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: v_mov_b32_e32 v0, s8
+; GFX908-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_add_i32 s10, s8, 0x400
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_pk_max_f16 v2, v1, v1
+; GFX908-NEXT: v_mov_b32_e32 v3, s10
+; GFX908-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v5, v0
+; GFX908-NEXT: v_pk_max_f16 v0, v5, v5
+; GFX908-NEXT: v_pk_max_f16 v4, v0, v2
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB12_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_mov_b32_e32 v0, s8
+; GFX8-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_add_i32 s10, s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_max_f16_sdwa v2, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v3, v1, v1
+; GFX8-NEXT: v_mov_b32_e32 v4, s10
+; GFX8-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_max_f16_sdwa v0, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v1, v6, v6
+; GFX8-NEXT: v_max_f16_sdwa v0, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v1, v1, v3
+; GFX8-NEXT: v_or_b32_e32 v5, v1, v0
+; GFX8-NEXT: v_mov_b32_e32 v0, v5
+; GFX8-NEXT: v_mov_b32_e32 v1, v6
+; GFX8-NEXT: buffer_atomic_cmpswap v[0:1], v4, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB12_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dword v3, v2, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v0
+; GFX7-NEXT: s_add_i32 s10, s8, 0x400
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, s10
+; GFX7-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v0, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v0
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v4, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB12_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dword v3, v2, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v0
+; GFX6-NEXT: s_add_i32 s10, s8, 0x400
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, s10
+; GFX6-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v6
+; GFX6-NEXT: v_or_b32_e32 v6, v0, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v5
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v0
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v4, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB12_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmax ptr addrspace(7) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define void @buffer_fat_ptr_agent_atomic_fmax_noret_v2f16__offset(ptr addrspace(7) inreg %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_v2f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v1, s4
+; GFX12-NEXT: s_addk_co_i32 s4, 0x400
+; GFX12-NEXT: v_pk_max_num_f16 v2, v0, v0
+; GFX12-NEXT: v_mov_b32_e32 v3, s4
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: buffer_load_b32 v1, v1, s[0:3], null offen offset:1024
+; GFX12-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v0, v1, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v0, v0, v2
+; GFX12-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v4
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB13_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_v2f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, s4
+; GFX940-NEXT: buffer_load_dword v1, v1, s[0:3], 0 offen offset:1024
+; GFX940-NEXT: s_add_i32 s6, s4, 0x400
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_pk_max_f16 v2, v0, v0
+; GFX940-NEXT: v_mov_b32_e32 v3, s6
+; GFX940-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v0, v1, v1
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_pk_max_f16 v0, v0, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[0:1]
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB13_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_v2f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v1, s4
+; GFX11-NEXT: s_addk_i32 s4, 0x400
+; GFX11-NEXT: v_pk_max_f16 v2, v0, v0
+; GFX11-NEXT: v_mov_b32_e32 v3, s4
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: buffer_load_b32 v1, v1, s[0:3], 0 offen offset:1024
+; GFX11-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v0, v1, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v0, v0, v2
+; GFX11-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v4
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB13_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_v2f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: v_pk_max_f16 v2, v0, v0
+; GFX10-NEXT: v_mov_b32_e32 v3, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v0, v1, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_pk_max_f16 v0, v0, v2
+; GFX10-NEXT: v_mov_b32_e32 v5, v1
+; GFX10-NEXT: v_mov_b32_e32 v4, v0
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v4
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB13_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_v2f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, s8
+; GFX90A-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX90A-NEXT: s_add_i32 s10, s8, 0x400
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_pk_max_f16 v2, v0, v0
+; GFX90A-NEXT: v_mov_b32_e32 v3, s10
+; GFX90A-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v0, v1, v1
+; GFX90A-NEXT: v_pk_max_f16 v0, v0, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB13_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_v2f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, s8
+; GFX908-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_add_i32 s10, s8, 0x400
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_pk_max_f16 v2, v0, v0
+; GFX908-NEXT: v_mov_b32_e32 v3, s10
+; GFX908-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v0, v1, v1
+; GFX908-NEXT: v_pk_max_f16 v0, v0, v2
+; GFX908-NEXT: v_mov_b32_e32 v5, v1
+; GFX908-NEXT: v_mov_b32_e32 v4, v0
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: v_mov_b32_e32 v1, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB13_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_v2f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, s8
+; GFX8-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_add_i32 s10, s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_max_f16_sdwa v2, v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v3, v0, v0
+; GFX8-NEXT: v_mov_b32_e32 v4, s10
+; GFX8-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v1, v1
+; GFX8-NEXT: v_max_f16_sdwa v0, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX8-NEXT: v_mov_b32_e32 v6, v1
+; GFX8-NEXT: v_mov_b32_e32 v5, v0
+; GFX8-NEXT: buffer_atomic_cmpswap v[5:6], v4, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v1
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v1, v5
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB13_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_v2f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dword v2, v2, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v0
+; GFX7-NEXT: s_add_i32 s10, s8, 0x400
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v5
+; GFX7-NEXT: v_mov_b32_e32 v2, s10
+; GFX7-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_max_f32_e32 v5, v5, v0
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_or_b32_e32 v5, v3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX7-NEXT: v_or_b32_e32 v4, v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v2, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB13_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_v2f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dword v2, v2, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v0
+; GFX6-NEXT: s_add_i32 s10, s8, 0x400
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v5
+; GFX6-NEXT: v_mov_b32_e32 v2, s10
+; GFX6-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_max_f32_e32 v5, v5, v0
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_or_b32_e32 v5, v3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX6-NEXT: v_or_b32_e32 v4, v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v2, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB13_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(7) %ptr, i32 256
+ %unused = atomicrmw fmax ptr addrspace(7) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x half> @buffer_fat_ptr_agent_atomic_fmax_ret_v2f16__offset__waterfall(ptr addrspace(7) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2f16__offset__waterfall:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX12-NEXT: s_mov_b32 s1, exec_lo
+; GFX12-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: buffer_load_b32 v6, v4, s[4:7], null offen offset:1024
+; GFX12-NEXT: ; implicit-def: $vgpr4
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB14_1
+; GFX12-NEXT: ; %bb.2:
+; GFX12-NEXT: s_mov_b32 exec_lo, s1
+; GFX12-NEXT: v_pk_max_num_f16 v8, v5, v5
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Loop Header: Depth=1
+; GFX12-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v4, v6, v6
+; GFX12-NEXT: s_mov_b32 s2, exec_lo
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v5, v4, v8
+; GFX12-NEXT: v_mov_b32_e32 v4, v5
+; GFX12-NEXT: v_mov_b32_e32 v5, v6
+; GFX12-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX12-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB14_4
+; GFX12-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX12-NEXT: s_mov_b32 exec_lo, s2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX12-NEXT: v_mov_b32_e32 v6, v4
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB14_3
+; GFX12-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2f16__offset__waterfall:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v8, 0x400, v4
+; GFX940-NEXT: s_mov_b64 s[2:3], exec
+; GFX940-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: buffer_load_dword v7, v4, s[4:7], 0 offen offset:1024
+; GFX940-NEXT: ; implicit-def: $vgpr4
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB14_1
+; GFX940-NEXT: ; %bb.2:
+; GFX940-NEXT: s_mov_b64 exec, s[2:3]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_pk_max_f16 v9, v5, v5
+; GFX940-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Loop Header: Depth=1
+; GFX940-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v4, v7, v7
+; GFX940-NEXT: s_mov_b64 s[8:9], exec
+; GFX940-NEXT: v_pk_max_f16 v6, v4, v9
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[6:7]
+; GFX940-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX940-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[4:7], 0 offen sc0
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB14_4
+; GFX940-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX940-NEXT: s_mov_b64 exec, s[8:9]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB14_3
+; GFX940-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2f16__offset__waterfall:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: buffer_load_b32 v6, v4, s[4:7], 0 offen offset:1024
+; GFX11-NEXT: ; implicit-def: $vgpr4
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB14_1
+; GFX11-NEXT: ; %bb.2:
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: v_pk_max_f16 v8, v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Loop Header: Depth=1
+; GFX11-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v4, v6, v6
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v5, v4, v8
+; GFX11-NEXT: v_mov_b32_e32 v4, v5
+; GFX11-NEXT: v_mov_b32_e32 v5, v6
+; GFX11-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX11-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], 0 offen glc
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB14_4
+; GFX11-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX11-NEXT: v_mov_b32_e32 v6, v4
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB14_3
+; GFX11-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2f16__offset__waterfall:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX10-NEXT: ; implicit-def: $vgpr4
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB14_1
+; GFX10-NEXT: ; %bb.2:
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: v_pk_max_f16 v8, v5, v5
+; GFX10-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Loop Header: Depth=1
+; GFX10-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v4, v6, v6
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_pk_max_f16 v5, v4, v8
+; GFX10-NEXT: v_mov_b32_e32 v4, v5
+; GFX10-NEXT: v_mov_b32_e32 v5, v6
+; GFX10-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX10-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB14_4
+; GFX10-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB14_3
+; GFX10-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2f16__offset__waterfall:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v8, 0x400, v4
+; GFX90A-NEXT: s_mov_b64 s[6:7], exec
+; GFX90A-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_nop 0
+; GFX90A-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen offset:1024
+; GFX90A-NEXT: ; implicit-def: $vgpr4
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_1
+; GFX90A-NEXT: ; %bb.2:
+; GFX90A-NEXT: s_mov_b64 exec, s[6:7]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_pk_max_f16 v9, v5, v5
+; GFX90A-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Loop Header: Depth=1
+; GFX90A-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v4, v7, v7
+; GFX90A-NEXT: v_pk_max_f16 v6, v4, v9
+; GFX90A-NEXT: s_mov_b64 s[12:13], exec
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[6:7] op_sel:[0,1]
+; GFX90A-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX90A-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[8:11], 0 offen glc
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_4
+; GFX90A-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX90A-NEXT: s_mov_b64 exec, s[12:13]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_3
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2f16__offset__waterfall:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v7, 0x400, v4
+; GFX908-NEXT: s_mov_b64 s[6:7], exec
+; GFX908-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_nop 0
+; GFX908-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX908-NEXT: ; implicit-def: $vgpr4
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_1
+; GFX908-NEXT: ; %bb.2:
+; GFX908-NEXT: s_mov_b64 exec, s[6:7]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_pk_max_f16 v8, v5, v5
+; GFX908-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Loop Header: Depth=1
+; GFX908-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v4, v6, v6
+; GFX908-NEXT: v_pk_max_f16 v5, v4, v8
+; GFX908-NEXT: v_mov_b32_e32 v4, v5
+; GFX908-NEXT: s_mov_b64 s[12:13], exec
+; GFX908-NEXT: v_mov_b32_e32 v5, v6
+; GFX908-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX908-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_4
+; GFX908-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX908-NEXT: s_mov_b64 exec, s[12:13]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_3
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2f16__offset__waterfall:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x400, v4
+; GFX8-NEXT: s_mov_b64 s[6:7], exec
+; GFX8-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_nop 0
+; GFX8-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX8-NEXT: ; implicit-def: $vgpr4
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_1
+; GFX8-NEXT: ; %bb.2:
+; GFX8-NEXT: s_mov_b64 exec, s[6:7]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_max_f16_sdwa v8, v5, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v9, v5, v5
+; GFX8-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Loop Header: Depth=1
+; GFX8-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v4, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v6, v6
+; GFX8-NEXT: v_max_f16_sdwa v4, v4, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_max_f16_e32 v5, v5, v9
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v4
+; GFX8-NEXT: v_mov_b32_e32 v4, v5
+; GFX8-NEXT: s_mov_b64 s[12:13], exec
+; GFX8-NEXT: v_mov_b32_e32 v5, v6
+; GFX8-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX8-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_4
+; GFX8-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX8-NEXT: s_mov_b64 exec, s[12:13]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_3
+; GFX8-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2f16__offset__waterfall:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v9, vcc, 0x400, v4
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen offset:1024
+; GFX7-NEXT: ; implicit-def: $vgpr4
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_1
+; GFX7-NEXT: ; %bb.2:
+; GFX7-NEXT: s_mov_b64 exec, s[6:7]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v5
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v10, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v11, v8
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Loop Header: Depth=1
+; GFX7-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: s_mov_b64 s[12:13], exec
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v10
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v11
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX7-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v9, s[8:11], 0 offen glc
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_4
+; GFX7-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX7-NEXT: s_mov_b64 exec, s[12:13]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_3
+; GFX7-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v0, v4
+; GFX7-NEXT: v_mov_b32_e32 v1, v5
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2f16__offset__waterfall:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v9, vcc, 0x400, v4
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen offset:1024
+; GFX6-NEXT: ; implicit-def: $vgpr4
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB14_1
+; GFX6-NEXT: ; %bb.2:
+; GFX6-NEXT: s_mov_b64 exec, s[6:7]
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v5
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v10, v6
+; GFX6-NEXT: v_cvt_f32_f16_e32 v11, v8
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Loop Header: Depth=1
+; GFX6-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: s_mov_b64 s[12:13], exec
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v10
+; GFX6-NEXT: v_max_f32_e32 v7, v7, v11
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX6-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX6-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v9, s[8:11], 0 offen glc
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB14_4
+; GFX6-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX6-NEXT: s_mov_b64 exec, s[12:13]
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB14_3
+; GFX6-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v0, v4
+; GFX6-NEXT: v_mov_b32_e32 v1, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmax ptr addrspace(7) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+; --------------------------------------------------------------------
+; <2 x bfloat>
+; --------------------------------------------------------------------
+
+define <2 x bfloat> @buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset(ptr addrspace(7) inreg %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s4
+; GFX12-NEXT: s_addk_co_i32 s4, 0x400
+; GFX12-NEXT: s_mov_b32 s5, 0
+; GFX12-NEXT: v_mov_b32_e32 v4, s4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX12-NEXT: buffer_load_b32 v0, v0, s[0:3], null offen offset:1024
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX12-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX12-NEXT: v_max_num_f32_e32 v1, v1, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v1, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX12-NEXT: v_add3_u32 v7, v7, v1, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_cndmask_b32 v1, v7, v9 :: v_dual_lshlrev_b32 v0, 16, v6
+; GFX12-NEXT: v_max_num_f32_e32 v0, v0, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX12-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX12-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v0, v5, v8, s4
+; GFX12-NEXT: v_perm_b32 v5, v1, v0, 0x7060302
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v0, v5 :: v_dual_mov_b32 v1, v6
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v4, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX12-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5
+; GFX12-NEXT: s_cbranch_execnz .LBB15_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, v0
+; GFX940-NEXT: v_mov_b32_e32 v0, s4
+; GFX940-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen offset:1024
+; GFX940-NEXT: s_addk_i32 s4, 0x400
+; GFX940-NEXT: s_mov_b64 s[6:7], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX940-NEXT: s_movk_i32 s8, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX940-NEXT: s_mov_b32 s9, 0x7060302
+; GFX940-NEXT: v_mov_b32_e32 v4, s4
+; GFX940-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v7
+; GFX940-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX940-NEXT: v_max_f32_e32 v1, v1, v3
+; GFX940-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v1, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v0
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX940-NEXT: v_add3_u32 v5, v5, v0, s8
+; GFX940-NEXT: v_add3_u32 v8, v8, v1, s8
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX940-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_cndmask_b32_e32 v1, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v0, v5, v6, s[4:5]
+; GFX940-NEXT: v_perm_b32 v6, v1, v0, s9
+; GFX940-NEXT: v_mov_b64_e32 v[0:1], v[6:7]
+; GFX940-NEXT: buffer_atomic_cmpswap v[0:1], v4, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v7
+; GFX940-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX940-NEXT: s_cbranch_execnz .LBB15_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s4
+; GFX11-NEXT: s_addk_i32 s4, 0x400
+; GFX11-NEXT: s_mov_b32 s5, 0
+; GFX11-NEXT: v_mov_b32_e32 v4, s4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX11-NEXT: buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX11-NEXT: v_max_f32_e32 v1, v1, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v1, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT: v_add3_u32 v7, v7, v1, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_cndmask_b32 v1, v7, v9 :: v_dual_lshlrev_b32 v0, 16, v6
+; GFX11-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX11-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX11-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v0, v5, v8, s4
+; GFX11-NEXT: v_perm_b32 v5, v1, v0, 0x7060302
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v0, v5 :: v_dual_mov_b32 v1, v6
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v4, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX11-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5
+; GFX11-NEXT: s_cbranch_execnz .LBB15_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_mov_b32_e32 v0, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: s_mov_b32 s9, 0
+; GFX10-NEXT: v_mov_b32_e32 v4, s8
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX10-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX10-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX10-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX10-NEXT: v_max_f32_e32 v1, v1, v3
+; GFX10-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v1, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX10-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v1, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s8, v0, v0
+; GFX10-NEXT: v_cndmask_b32_e32 v1, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v0, v5, v8, s8
+; GFX10-NEXT: v_perm_b32 v5, v1, v0, 0x7060302
+; GFX10-NEXT: v_mov_b32_e32 v0, v5
+; GFX10-NEXT: v_mov_b32_e32 v1, v6
+; GFX10-NEXT: buffer_atomic_cmpswap v[0:1], v4, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s9, vcc_lo, s9
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s9
+; GFX10-NEXT: s_cbranch_execnz .LBB15_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s9
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: v_mov_b32_e32 v0, s8
+; GFX90A-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX90A-NEXT: s_addk_i32 s8, 0x400
+; GFX90A-NEXT: s_mov_b64 s[10:11], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX90A-NEXT: s_movk_i32 s12, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX90A-NEXT: s_mov_b32 s13, 0x7060302
+; GFX90A-NEXT: v_mov_b32_e32 v4, s8
+; GFX90A-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v7
+; GFX90A-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX90A-NEXT: v_max_f32_e32 v1, v1, v3
+; GFX90A-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v1, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v0
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX90A-NEXT: v_add3_u32 v5, v5, v0, s12
+; GFX90A-NEXT: v_add3_u32 v8, v8, v1, s12
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[8:9], v0, v0
+; GFX90A-NEXT: v_cndmask_b32_e64 v0, v5, v6, s[8:9]
+; GFX90A-NEXT: v_cndmask_b32_e32 v1, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v1, v0, s13
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[6:7] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[0:1], v4, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v7
+; GFX90A-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX90A-NEXT: s_cbranch_execnz .LBB15_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: v_mov_b32_e32 v0, s8
+; GFX908-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_addk_i32 s8, 0x400
+; GFX908-NEXT: s_mov_b64 s[10:11], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX908-NEXT: s_movk_i32 s12, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX908-NEXT: s_mov_b32 s13, 0x7060302
+; GFX908-NEXT: v_mov_b32_e32 v4, s8
+; GFX908-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX908-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX908-NEXT: v_max_f32_e32 v1, v1, v3
+; GFX908-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v1, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX908-NEXT: v_add3_u32 v5, v5, v0, s12
+; GFX908-NEXT: v_add3_u32 v8, v8, v1, s12
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX908-NEXT: v_cmp_u_f32_e64 s[8:9], v0, v0
+; GFX908-NEXT: v_cndmask_b32_e64 v0, v5, v7, s[8:9]
+; GFX908-NEXT: v_cndmask_b32_e32 v1, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v5, v1, v0, s13
+; GFX908-NEXT: v_mov_b32_e32 v0, v5
+; GFX908-NEXT: v_mov_b32_e32 v1, v6
+; GFX908-NEXT: buffer_atomic_cmpswap v[0:1], v4, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX908-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX908-NEXT: s_cbranch_execnz .LBB15_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_mov_b32_e32 v0, s8
+; GFX8-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_addk_i32 s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[10:11], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX8-NEXT: v_mov_b32_e32 v4, s8
+; GFX8-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX8-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX8-NEXT: v_max_f32_e32 v1, v1, v3
+; GFX8-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v1, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v0
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[8:9], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v1, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v5, v7, s[8:9]
+; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX8-NEXT: v_alignbit_b32 v5, v1, v0, 16
+; GFX8-NEXT: v_mov_b32_e32 v0, v5
+; GFX8-NEXT: v_mov_b32_e32 v1, v6
+; GFX8-NEXT: buffer_atomic_cmpswap v[0:1], v4, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX8-NEXT: s_cbranch_execnz .LBB15_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dword v4, v2, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: s_addk_i32 s8, 0x400
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: s_mov_b64 s[10:11], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, s8
+; GFX7-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v3
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v5
+; GFX7-NEXT: v_alignbit_b32 v0, v0, v6, 16
+; GFX7-NEXT: v_mov_b32_e32 v6, v1
+; GFX7-NEXT: v_mov_b32_e32 v5, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v4, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v5
+; GFX7-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX7-NEXT: s_cbranch_execnz .LBB15_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dword v4, v2, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: s_addk_i32 s8, 0x400
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: s_mov_b64 s[10:11], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v1
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, s8
+; GFX6-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v0
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_max_f32_e32 v5, v5, v2
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v3
+; GFX6-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 16, v5
+; GFX6-NEXT: v_alignbit_b32 v0, v0, v6, 16
+; GFX6-NEXT: v_mov_b32_e32 v6, v1
+; GFX6-NEXT: v_mov_b32_e32 v5, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v4, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v1
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v5
+; GFX6-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX6-NEXT: s_cbranch_execnz .LBB15_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmax ptr addrspace(7) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @buffer_fat_ptr_agent_atomic_fmax_noret_v2bf16__offset(ptr addrspace(7) inreg %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_v2bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v1, s4 :: v_dual_lshlrev_b32 v2, 16, v0
+; GFX12-NEXT: s_addk_co_i32 s4, 0x400
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_and_b32 v3, 0xffff0000, v0
+; GFX12-NEXT: buffer_load_b32 v1, v1, s[0:3], null offen offset:1024
+; GFX12-NEXT: s_mov_b32 s5, 0
+; GFX12-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_max_num_f32 v5, v5, v3 :: v_dual_max_num_f32 v0, v0, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_add3_u32 v6, v6, v0, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v0, v6, v8, s4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v0, v5, v0, 0x7060302
+; GFX12-NEXT: v_dual_mov_b32 v6, v1 :: v_dual_mov_b32 v5, v0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[5:6], v4, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v5
+; GFX12-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5
+; GFX12-NEXT: s_cbranch_execnz .LBB16_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_v2bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, s4
+; GFX940-NEXT: buffer_load_dword v1, v1, s[0:3], 0 offen offset:1024
+; GFX940-NEXT: s_addk_i32 s4, 0x400
+; GFX940-NEXT: s_mov_b64 s[6:7], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v0
+; GFX940-NEXT: s_movk_i32 s8, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX940-NEXT: s_mov_b32 s9, 0x7060302
+; GFX940-NEXT: v_mov_b32_e32 v4, s4
+; GFX940-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX940-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v3
+; GFX940-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v0, s8
+; GFX940-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v0, v6, v7, s[4:5]
+; GFX940-NEXT: v_perm_b32 v0, v5, v0, s9
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[0:1]
+; GFX940-NEXT: buffer_atomic_cmpswap v[6:7], v4, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v6, v1
+; GFX940-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX940-NEXT: v_mov_b32_e32 v1, v6
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX940-NEXT: s_cbranch_execnz .LBB16_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_v2bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v1, s4 :: v_dual_lshlrev_b32 v2, 16, v0
+; GFX11-NEXT: s_addk_i32 s4, 0x400
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_and_b32 v3, 0xffff0000, v0
+; GFX11-NEXT: buffer_load_b32 v1, v1, s[0:3], 0 offen offset:1024
+; GFX11-NEXT: s_mov_b32 s5, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_max_f32 v5, v5, v3 :: v_dual_max_f32 v0, v0, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_add3_u32 v6, v6, v0, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v0, v6, v8, s4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v0, v5, v0, 0x7060302
+; GFX11-NEXT: v_dual_mov_b32 v6, v1 :: v_dual_mov_b32 v5, v0
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[5:6], v4, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v5
+; GFX11-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5
+; GFX11-NEXT: s_cbranch_execnz .LBB16_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_v2bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v0
+; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX10-NEXT: v_mov_b32_e32 v4, s8
+; GFX10-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: s_mov_b32 s9, 0
+; GFX10-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v3
+; GFX10-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v6, v6, v0, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s8, v0, v0
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v0, v6, v8, s8
+; GFX10-NEXT: v_perm_b32 v0, v5, v0, 0x7060302
+; GFX10-NEXT: v_mov_b32_e32 v6, v1
+; GFX10-NEXT: v_mov_b32_e32 v5, v0
+; GFX10-NEXT: buffer_atomic_cmpswap v[5:6], v4, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: s_or_b32 s9, vcc_lo, s9
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s9
+; GFX10-NEXT: s_cbranch_execnz .LBB16_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s9
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_v2bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, s8
+; GFX90A-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX90A-NEXT: s_addk_i32 s8, 0x400
+; GFX90A-NEXT: s_mov_b64 s[10:11], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v0
+; GFX90A-NEXT: s_movk_i32 s12, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX90A-NEXT: s_mov_b32 s13, 0x7060302
+; GFX90A-NEXT: v_mov_b32_e32 v4, s8
+; GFX90A-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX90A-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v3
+; GFX90A-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v0, s12
+; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s12
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[8:9], v0, v0
+; GFX90A-NEXT: v_cndmask_b32_e64 v0, v6, v7, s[8:9]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v0, v5, v0, s13
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[6:7], v4, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v6, v1
+; GFX90A-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v6
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX90A-NEXT: s_cbranch_execnz .LBB16_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_v2bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, s8
+; GFX908-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_addk_i32 s8, 0x400
+; GFX908-NEXT: s_mov_b64 s[10:11], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v0
+; GFX908-NEXT: s_movk_i32 s12, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX908-NEXT: s_mov_b32 s13, 0x7060302
+; GFX908-NEXT: v_mov_b32_e32 v4, s8
+; GFX908-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX908-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v3
+; GFX908-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v6, v6, v0, s12
+; GFX908-NEXT: v_add3_u32 v8, v8, v5, s12
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[8:9], v0, v0
+; GFX908-NEXT: v_cndmask_b32_e64 v0, v6, v7, s[8:9]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v0, v5, v0, s13
+; GFX908-NEXT: v_mov_b32_e32 v6, v1
+; GFX908-NEXT: v_mov_b32_e32 v5, v0
+; GFX908-NEXT: buffer_atomic_cmpswap v[5:6], v4, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v1
+; GFX908-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX908-NEXT: s_cbranch_execnz .LBB16_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_v2bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, s8
+; GFX8-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_addk_i32 s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[10:11], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v0
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX8-NEXT: v_mov_b32_e32 v4, s8
+; GFX8-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX8-NEXT: v_max_f32_e32 v0, v0, v2
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v3
+; GFX8-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v0
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[8:9], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v6, v7, s[8:9]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v0, v5, v0, 16
+; GFX8-NEXT: v_mov_b32_e32 v6, v1
+; GFX8-NEXT: v_mov_b32_e32 v5, v0
+; GFX8-NEXT: buffer_atomic_cmpswap v[5:6], v4, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v1
+; GFX8-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX8-NEXT: v_mov_b32_e32 v1, v5
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX8-NEXT: s_cbranch_execnz .LBB16_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_v2bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dword v2, v2, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: s_addk_i32 s8, 0x400
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v0
+; GFX7-NEXT: s_mov_b64 s[10:11], 0
+; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v2
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_max_f32_e32 v5, v5, v0
+; GFX7-NEXT: v_max_f32_e32 v6, v6, v1
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_alignbit_b32 v3, v3, v6, 16
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v2, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX7-NEXT: s_cbranch_execnz .LBB16_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmax_noret_v2bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dword v2, v2, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: s_addk_i32 s8, 0x400
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v0
+; GFX6-NEXT: s_mov_b64 s[10:11], 0
+; GFX6-NEXT: v_and_b32_e32 v0, 0xffff0000, v1
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v2
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_max_f32_e32 v5, v5, v0
+; GFX6-NEXT: v_max_f32_e32 v6, v6, v1
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_alignbit_b32 v3, v3, v6, 16
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v2, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX6-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX6-NEXT: s_cbranch_execnz .LBB16_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(7) %ptr, i32 256
+ %unused = atomicrmw fmax ptr addrspace(7) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x bfloat> @buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset__waterfall(ptr addrspace(7) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset__waterfall:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX12-NEXT: s_mov_b32 s1, exec_lo
+; GFX12-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: buffer_load_b32 v6, v4, s[4:7], null offen offset:1024
+; GFX12-NEXT: ; implicit-def: $vgpr4
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB17_1
+; GFX12-NEXT: ; %bb.2:
+; GFX12-NEXT: s_mov_b32 exec_lo, s1
+; GFX12-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX12-NEXT: v_and_b32_e32 v9, 0xffff0000, v5
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Loop Header: Depth=1
+; GFX12-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX12-NEXT: s_mov_b32 s2, exec_lo
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_max_num_f32 v5, v5, v9 :: v_dual_max_num_f32 v4, v4, v8
+; GFX12-NEXT: v_bfe_u32 v11, v5, 16, 1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v10, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v12, 0x400000, v4
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX12-NEXT: v_or_b32_e32 v13, 0x400000, v5
+; GFX12-NEXT: v_add3_u32 v11, v11, v5, 0x7fff
+; GFX12-NEXT: v_add3_u32 v10, v10, v4, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v4, v10, v12, vcc_lo
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v11, v13, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v5, v5, v4, 0x7060302
+; GFX12-NEXT: v_mov_b32_e32 v4, v5
+; GFX12-NEXT: v_mov_b32_e32 v5, v6
+; GFX12-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX12-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB17_4
+; GFX12-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX12-NEXT: s_mov_b32 exec_lo, s2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX12-NEXT: v_mov_b32_e32 v6, v4
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB17_3
+; GFX12-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset__waterfall:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v8, 0x400, v4
+; GFX940-NEXT: s_mov_b64 s[2:3], exec
+; GFX940-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: buffer_load_dword v7, v4, s[4:7], 0 offen offset:1024
+; GFX940-NEXT: ; implicit-def: $vgpr4
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB17_1
+; GFX940-NEXT: ; %bb.2:
+; GFX940-NEXT: s_mov_b64 exec, s[2:3]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v9, 16, v5
+; GFX940-NEXT: s_movk_i32 s10, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v10, 0xffff0000, v5
+; GFX940-NEXT: s_mov_b32 s11, 0x7060302
+; GFX940-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Loop Header: Depth=1
+; GFX940-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v7
+; GFX940-NEXT: v_max_f32_e32 v4, v4, v9
+; GFX940-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX940-NEXT: v_add3_u32 v5, v5, v4, s10
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_mov_b64 s[8:9], exec
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_max_f32_e32 v5, v5, v10
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s10
+; GFX940-NEXT: v_or_b32_e32 v11, 0x400000, v5
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v11, vcc
+; GFX940-NEXT: v_perm_b32 v6, v5, v4, s11
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[6:7]
+; GFX940-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX940-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[4:7], 0 offen sc0
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB17_4
+; GFX940-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX940-NEXT: s_mov_b64 exec, s[8:9]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB17_3
+; GFX940-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset__waterfall:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: buffer_load_b32 v6, v4, s[4:7], 0 offen offset:1024
+; GFX11-NEXT: ; implicit-def: $vgpr4
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB17_1
+; GFX11-NEXT: ; %bb.2:
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v5
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Loop Header: Depth=1
+; GFX11-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_max_f32 v5, v5, v9 :: v_dual_max_f32 v4, v4, v8
+; GFX11-NEXT: v_bfe_u32 v11, v5, 16, 1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v10, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v4
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v5
+; GFX11-NEXT: v_add3_u32 v11, v11, v5, 0x7fff
+; GFX11-NEXT: v_add3_u32 v10, v10, v4, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v4, v10, v12, vcc_lo
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v11, v13, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v5, v5, v4, 0x7060302
+; GFX11-NEXT: v_mov_b32_e32 v4, v5
+; GFX11-NEXT: v_mov_b32_e32 v5, v6
+; GFX11-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX11-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], 0 offen glc
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB17_4
+; GFX11-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX11-NEXT: v_mov_b32_e32 v6, v4
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB17_3
+; GFX11-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset__waterfall:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX10-NEXT: ; implicit-def: $vgpr4
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB17_1
+; GFX10-NEXT: ; %bb.2:
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX10-NEXT: v_and_b32_e32 v9, 0xffff0000, v5
+; GFX10-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Loop Header: Depth=1
+; GFX10-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_max_f32_e32 v4, v4, v8
+; GFX10-NEXT: v_max_f32_e32 v5, v5, v9
+; GFX10-NEXT: v_bfe_u32 v10, v4, 16, 1
+; GFX10-NEXT: v_bfe_u32 v11, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v12, 0x400000, v4
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX10-NEXT: v_or_b32_e32 v13, 0x400000, v5
+; GFX10-NEXT: v_add3_u32 v10, v10, v4, 0x7fff
+; GFX10-NEXT: v_add3_u32 v11, v11, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v4, v10, v12, vcc_lo
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v11, v13, vcc_lo
+; GFX10-NEXT: v_perm_b32 v5, v5, v4, 0x7060302
+; GFX10-NEXT: v_mov_b32_e32 v4, v5
+; GFX10-NEXT: v_mov_b32_e32 v5, v6
+; GFX10-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX10-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB17_4
+; GFX10-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB17_3
+; GFX10-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset__waterfall:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v8, 0x400, v4
+; GFX90A-NEXT: s_mov_b64 s[6:7], exec
+; GFX90A-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_nop 0
+; GFX90A-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen offset:1024
+; GFX90A-NEXT: ; implicit-def: $vgpr4
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB17_1
+; GFX90A-NEXT: ; %bb.2:
+; GFX90A-NEXT: s_mov_b64 exec, s[6:7]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v9, 16, v5
+; GFX90A-NEXT: s_movk_i32 s14, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v10, 0xffff0000, v5
+; GFX90A-NEXT: s_mov_b32 s15, 0x7060302
+; GFX90A-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Loop Header: Depth=1
+; GFX90A-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v7
+; GFX90A-NEXT: v_max_f32_e32 v4, v4, v9
+; GFX90A-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX90A-NEXT: v_add3_u32 v5, v5, v4, s14
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_max_f32_e32 v5, v5, v10
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s14
+; GFX90A-NEXT: v_or_b32_e32 v11, 0x400000, v5
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v11, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v4, s15
+; GFX90A-NEXT: s_mov_b64 s[12:13], exec
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[6:7] op_sel:[0,1]
+; GFX90A-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX90A-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[8:11], 0 offen glc
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB17_4
+; GFX90A-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX90A-NEXT: s_mov_b64 exec, s[12:13]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB17_3
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset__waterfall:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v7, 0x400, v4
+; GFX908-NEXT: s_mov_b64 s[6:7], exec
+; GFX908-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_nop 0
+; GFX908-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX908-NEXT: ; implicit-def: $vgpr4
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_1
+; GFX908-NEXT: ; %bb.2:
+; GFX908-NEXT: s_mov_b64 exec, s[6:7]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX908-NEXT: s_movk_i32 s14, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v9, 0xffff0000, v5
+; GFX908-NEXT: s_mov_b32 s15, 0x7060302
+; GFX908-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Loop Header: Depth=1
+; GFX908-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX908-NEXT: v_max_f32_e32 v4, v4, v8
+; GFX908-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX908-NEXT: v_add3_u32 v5, v5, v4, s14
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v4
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX908-NEXT: v_cndmask_b32_e32 v4, v5, v10, vcc
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_max_f32_e32 v5, v5, v9
+; GFX908-NEXT: v_bfe_u32 v10, v5, 16, 1
+; GFX908-NEXT: v_add3_u32 v10, v10, v5, s14
+; GFX908-NEXT: v_or_b32_e32 v11, 0x400000, v5
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v10, v11, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v4, s15
+; GFX908-NEXT: v_mov_b32_e32 v4, v5
+; GFX908-NEXT: s_mov_b64 s[12:13], exec
+; GFX908-NEXT: v_mov_b32_e32 v5, v6
+; GFX908-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX908-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_4
+; GFX908-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX908-NEXT: s_mov_b64 exec, s[12:13]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_3
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset__waterfall:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x400, v4
+; GFX8-NEXT: s_mov_b64 s[6:7], exec
+; GFX8-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_nop 0
+; GFX8-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX8-NEXT: ; implicit-def: $vgpr4
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_1
+; GFX8-NEXT: ; %bb.2:
+; GFX8-NEXT: s_mov_b64 exec, s[6:7]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX8-NEXT: v_and_b32_e32 v9, 0xffff0000, v5
+; GFX8-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Loop Header: Depth=1
+; GFX8-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX8-NEXT: v_max_f32_e32 v4, v4, v8
+; GFX8-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v4
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v4, v5, v10, vcc
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_max_f32_e32 v5, v5, v9
+; GFX8-NEXT: v_bfe_u32 v10, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v10, vcc, v10, v5
+; GFX8-NEXT: v_add_u32_e32 v10, vcc, 0x7fff, v10
+; GFX8-NEXT: v_or_b32_e32 v11, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v10, v11, vcc
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX8-NEXT: v_mov_b32_e32 v4, v5
+; GFX8-NEXT: s_mov_b64 s[12:13], exec
+; GFX8-NEXT: v_mov_b32_e32 v5, v6
+; GFX8-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX8-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_4
+; GFX8-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX8-NEXT: s_mov_b64 exec, s[12:13]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_3
+; GFX8-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset__waterfall:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v8, vcc, 0x400, v4
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen offset:1024
+; GFX7-NEXT: ; implicit-def: $vgpr4
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_1
+; GFX7-NEXT: ; %bb.2:
+; GFX7-NEXT: s_mov_b64 exec, s[6:7]
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v7
+; GFX7-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v6
+; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v5
+; GFX7-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Loop Header: Depth=1
+; GFX7-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v7
+; GFX7-NEXT: v_max_f32_e32 v4, v4, v9
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_max_f32_e32 v7, v7, v10
+; GFX7-NEXT: v_alignbit_b32 v5, v5, v6, 16
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: s_mov_b64 s[12:13], exec
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX7-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v8, s[8:11], 0 offen glc
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_4
+; GFX7-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX7-NEXT: s_mov_b64 exec, s[12:13]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: v_lshlrev_b32_e32 v7, 16, v6
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_3
+; GFX7-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v0, v7
+; GFX7-NEXT: v_mov_b32_e32 v1, v4
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_v2bf16__offset__waterfall:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v8, vcc, 0x400, v4
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen offset:1024
+; GFX6-NEXT: ; implicit-def: $vgpr4
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB17_1
+; GFX6-NEXT: ; %bb.2:
+; GFX6-NEXT: s_mov_b64 exec, s[6:7]
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v6
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v7
+; GFX6-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: v_and_b32_e32 v9, 0xffff0000, v6
+; GFX6-NEXT: v_and_b32_e32 v10, 0xffff0000, v5
+; GFX6-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Loop Header: Depth=1
+; GFX6-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v4
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v7
+; GFX6-NEXT: v_max_f32_e32 v4, v4, v9
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v6
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_max_f32_e32 v7, v7, v10
+; GFX6-NEXT: v_alignbit_b32 v5, v5, v6, 16
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: s_mov_b64 s[12:13], exec
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX6-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v8, s[8:11], 0 offen glc
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB17_4
+; GFX6-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX6-NEXT: s_mov_b64 exec, s[12:13]
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_lshlrev_b32_e32 v7, 16, v6
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB17_3
+; GFX6-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v0, v7
+; GFX6-NEXT: v_mov_b32_e32 v1, v4
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmax ptr addrspace(7) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+; --------------------------------------------------------------------
+; misc
+; --------------------------------------------------------------------
+
+define float @buffer_fat_ptr_system_atomic_fmax_ret_f32__offset(ptr addrspace(7) inreg %ptr, float %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_system_atomic_fmax_ret_f32__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s4
+; GFX12-NEXT: s_addk_co_i32 s4, 0x400
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: v_dual_mov_b32 v3, s4 :: v_dual_max_num_f32 v2, v1, v1
+; GFX12-NEXT: buffer_load_b32 v0, v0, s[0:3], null offen offset:1024
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v5, v0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v0, v5, v5
+; GFX12-NEXT: v_max_num_f32_e32 v4, v0, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v3, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB18_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_system_atomic_fmax_ret_f32__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, v0
+; GFX940-NEXT: v_mov_b32_e32 v0, s4
+; GFX940-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen offset:1024
+; GFX940-NEXT: s_add_i32 s6, s4, 0x400
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX940-NEXT: v_mov_b32_e32 v3, s6
+; GFX940-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v0
+; GFX940-NEXT: v_max_f32_e32 v0, v5, v5
+; GFX940-NEXT: v_max_f32_e32 v4, v0, v2
+; GFX940-NEXT: v_mov_b64_e32 v[0:1], v[4:5]
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB18_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_system_atomic_fmax_ret_f32__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s4
+; GFX11-NEXT: s_addk_i32 s4, 0x400
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: v_dual_mov_b32 v3, s4 :: v_dual_max_f32 v2, v1, v1
+; GFX11-NEXT: buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v5, v0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v0, v5, v5
+; GFX11-NEXT: v_max_f32_e32 v4, v0, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v3, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB18_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_system_atomic_fmax_ret_f32__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_mov_b32_e32 v0, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: v_mov_b32_e32 v3, s8
+; GFX10-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX10-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v5, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_max_f32_e32 v0, v5, v5
+; GFX10-NEXT: v_max_f32_e32 v4, v0, v2
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB18_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_system_atomic_fmax_ret_f32__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: v_mov_b32_e32 v0, s8
+; GFX90A-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX90A-NEXT: s_add_i32 s10, s8, 0x400
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX90A-NEXT: v_mov_b32_e32 v3, s10
+; GFX90A-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v0
+; GFX90A-NEXT: v_max_f32_e32 v0, v5, v5
+; GFX90A-NEXT: v_max_f32_e32 v4, v0, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB18_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_system_atomic_fmax_ret_f32__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: v_mov_b32_e32 v0, s8
+; GFX908-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_add_i32 s10, s8, 0x400
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX908-NEXT: v_mov_b32_e32 v3, s10
+; GFX908-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v5, v0
+; GFX908-NEXT: v_max_f32_e32 v0, v5, v5
+; GFX908-NEXT: v_max_f32_e32 v4, v0, v2
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB18_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_system_atomic_fmax_ret_f32__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_mov_b32_e32 v0, s8
+; GFX8-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_add_i32 s10, s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX8-NEXT: v_mov_b32_e32 v3, s10
+; GFX8-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v5, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v5
+; GFX8-NEXT: v_max_f32_e32 v4, v0, v2
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: v_mov_b32_e32 v1, v5
+; GFX8-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB18_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_system_atomic_fmax_ret_f32__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v1, v0
+; GFX7-NEXT: v_mov_b32_e32 v0, s8
+; GFX7-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: s_add_i32 s10, s8, 0x400
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, s10
+; GFX7-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v5
+; GFX7-NEXT: v_max_f32_e32 v4, v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v0, v4
+; GFX7-NEXT: v_mov_b32_e32 v1, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB18_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_system_atomic_fmax_ret_f32__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v1, v0
+; GFX6-NEXT: v_mov_b32_e32 v0, s8
+; GFX6-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: s_add_i32 s10, s8, 0x400
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, s10
+; GFX6-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v5
+; GFX6-NEXT: v_max_f32_e32 v4, v0, v2
+; GFX6-NEXT: v_mov_b32_e32 v0, v4
+; GFX6-NEXT: v_mov_b32_e32 v1, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB18_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmax ptr addrspace(7) %gep, float %val seq_cst
+ ret float %result
+}
+
+attributes #0 = { nounwind "amdgpu-unsafe-fp-atomics"="true" }
+
+
diff --git a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmin.ll b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmin.ll
new file mode 100644
index 0000000000000..51f63c93af57b
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmin.ll
@@ -0,0 +1,8062 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 < %s | FileCheck -check-prefix=GFX12 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx940 < %s | FileCheck -check-prefix=GFX940 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx90a < %s | FileCheck -check-prefix=GFX90A %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx908 < %s | FileCheck -check-prefix=GFX908 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tonga < %s | FileCheck -check-prefix=GFX8 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=hawaii < %s | FileCheck -check-prefix=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tahiti < %s | FileCheck -check-prefix=GFX6 %s
+
+; --------------------------------------------------------------------
+; float
+; --------------------------------------------------------------------
+
+define float @buffer_fat_ptr_agent_atomic_fmin_ret_f32__offset(ptr addrspace(7) inreg %ptr, float %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f32__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s4
+; GFX12-NEXT: s_addk_co_i32 s4, 0x400
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: v_dual_mov_b32 v3, s4 :: v_dual_max_num_f32 v2, v1, v1
+; GFX12-NEXT: buffer_load_b32 v0, v0, s[0:3], null offen offset:1024
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v5, v0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v0, v5, v5
+; GFX12-NEXT: v_min_num_f32_e32 v4, v0, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v3, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB0_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f32__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, v0
+; GFX940-NEXT: v_mov_b32_e32 v0, s4
+; GFX940-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen offset:1024
+; GFX940-NEXT: s_add_i32 s6, s4, 0x400
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX940-NEXT: v_mov_b32_e32 v3, s6
+; GFX940-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v0
+; GFX940-NEXT: v_max_f32_e32 v0, v5, v5
+; GFX940-NEXT: v_min_f32_e32 v4, v0, v2
+; GFX940-NEXT: v_mov_b64_e32 v[0:1], v[4:5]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB0_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f32__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s4
+; GFX11-NEXT: s_addk_i32 s4, 0x400
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: v_dual_mov_b32 v3, s4 :: v_dual_max_f32 v2, v1, v1
+; GFX11-NEXT: buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v5, v0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v0, v5, v5
+; GFX11-NEXT: v_min_f32_e32 v4, v0, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v3, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB0_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f32__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_mov_b32_e32 v0, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: v_mov_b32_e32 v3, s8
+; GFX10-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX10-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v5, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_max_f32_e32 v0, v5, v5
+; GFX10-NEXT: v_min_f32_e32 v4, v0, v2
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB0_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f32__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: v_mov_b32_e32 v0, s8
+; GFX90A-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX90A-NEXT: s_add_i32 s10, s8, 0x400
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX90A-NEXT: v_mov_b32_e32 v3, s10
+; GFX90A-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v0
+; GFX90A-NEXT: v_max_f32_e32 v0, v5, v5
+; GFX90A-NEXT: v_min_f32_e32 v4, v0, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB0_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f32__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: v_mov_b32_e32 v0, s8
+; GFX908-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_add_i32 s10, s8, 0x400
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX908-NEXT: v_mov_b32_e32 v3, s10
+; GFX908-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v5, v0
+; GFX908-NEXT: v_max_f32_e32 v0, v5, v5
+; GFX908-NEXT: v_min_f32_e32 v4, v0, v2
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB0_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f32__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_mov_b32_e32 v0, s8
+; GFX8-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_add_i32 s10, s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX8-NEXT: v_mov_b32_e32 v3, s10
+; GFX8-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v5, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v5
+; GFX8-NEXT: v_min_f32_e32 v4, v0, v2
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: v_mov_b32_e32 v1, v5
+; GFX8-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB0_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f32__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v1, v0
+; GFX7-NEXT: v_mov_b32_e32 v0, s8
+; GFX7-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: s_add_i32 s10, s8, 0x400
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, s10
+; GFX7-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v5
+; GFX7-NEXT: v_min_f32_e32 v4, v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v0, v4
+; GFX7-NEXT: v_mov_b32_e32 v1, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB0_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f32__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v1, v0
+; GFX6-NEXT: v_mov_b32_e32 v0, s8
+; GFX6-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: s_add_i32 s10, s8, 0x400
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, s10
+; GFX6-NEXT: .LBB0_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v5
+; GFX6-NEXT: v_min_f32_e32 v4, v0, v2
+; GFX6-NEXT: v_mov_b32_e32 v0, v4
+; GFX6-NEXT: v_mov_b32_e32 v1, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB0_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmin ptr addrspace(7) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+define void @buffer_fat_ptr_agent_atomic_fmin_noret_f32__offset(ptr addrspace(7) inreg %ptr, float %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f32__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v1, s4 :: v_dual_max_num_f32 v2, v0, v0
+; GFX12-NEXT: s_addk_co_i32 s4, 0x400
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_mov_b32_e32 v3, s4
+; GFX12-NEXT: buffer_load_b32 v1, v1, s[0:3], null offen offset:1024
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v0, v1, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v0, v0, v2
+; GFX12-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v4
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB1_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f32__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, s4
+; GFX940-NEXT: buffer_load_dword v1, v1, s[0:3], 0 offen offset:1024
+; GFX940-NEXT: s_add_i32 s6, s4, 0x400
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v0, v0
+; GFX940-NEXT: v_mov_b32_e32 v3, s6
+; GFX940-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v0, v1, v1
+; GFX940-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[0:1]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB1_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f32__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v1, s4 :: v_dual_max_f32 v2, v0, v0
+; GFX11-NEXT: s_addk_i32 s4, 0x400
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v3, s4
+; GFX11-NEXT: buffer_load_b32 v1, v1, s[0:3], 0 offen offset:1024
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v0, v1, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX11-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v4
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB1_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f32__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: v_max_f32_e32 v2, v0, v0
+; GFX10-NEXT: v_mov_b32_e32 v3, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v0, v1, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX10-NEXT: v_mov_b32_e32 v5, v1
+; GFX10-NEXT: v_mov_b32_e32 v4, v0
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v4
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB1_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f32__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, s8
+; GFX90A-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX90A-NEXT: s_add_i32 s10, s8, 0x400
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v0, v0
+; GFX90A-NEXT: v_mov_b32_e32 v3, s10
+; GFX90A-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v0, v1, v1
+; GFX90A-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB1_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f32__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, s8
+; GFX908-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_add_i32 s10, s8, 0x400
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v0, v0
+; GFX908-NEXT: v_mov_b32_e32 v3, s10
+; GFX908-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v0, v1, v1
+; GFX908-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX908-NEXT: v_mov_b32_e32 v5, v1
+; GFX908-NEXT: v_mov_b32_e32 v4, v0
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: v_mov_b32_e32 v1, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB1_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f32__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, s8
+; GFX8-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_add_i32 s10, s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v0
+; GFX8-NEXT: v_mov_b32_e32 v3, s10
+; GFX8-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v1
+; GFX8-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX8-NEXT: v_mov_b32_e32 v5, v1
+; GFX8-NEXT: v_mov_b32_e32 v4, v0
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v1, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB1_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f32__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v1, s8
+; GFX7-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: s_add_i32 s10, s8, 0x400
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v0
+; GFX7-NEXT: v_mov_b32_e32 v3, s10
+; GFX7-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v1
+; GFX7-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v5, v1
+; GFX7-NEXT: v_mov_b32_e32 v4, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v1, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB1_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f32__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v1, s8
+; GFX6-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: s_add_i32 s10, s8, 0x400
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v0
+; GFX6-NEXT: v_mov_b32_e32 v3, s10
+; GFX6-NEXT: .LBB1_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v1
+; GFX6-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v1
+; GFX6-NEXT: v_mov_b32_e32 v4, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v1, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB1_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(7) %ptr, i32 256
+ %unused = atomicrmw fmin ptr addrspace(7) %gep, float %val syncscope("agent") seq_cst
+ ret void
+}
+
+define float @buffer_fat_ptr_agent_atomic_fmin_ret_f32__offset__waterfall(ptr addrspace(7) %ptr, float %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f32__offset__waterfall:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX12-NEXT: s_mov_b32 s1, exec_lo
+; GFX12-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: buffer_load_b32 v6, v4, s[4:7], null offen offset:1024
+; GFX12-NEXT: ; implicit-def: $vgpr4
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB2_1
+; GFX12-NEXT: ; %bb.2:
+; GFX12-NEXT: s_mov_b32 exec_lo, s1
+; GFX12-NEXT: v_max_num_f32_e32 v8, v5, v5
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Loop Header: Depth=1
+; GFX12-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f32_e32 v4, v6, v6
+; GFX12-NEXT: s_mov_b32 s2, exec_lo
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f32_e32 v5, v4, v8
+; GFX12-NEXT: v_mov_b32_e32 v4, v5
+; GFX12-NEXT: v_mov_b32_e32 v5, v6
+; GFX12-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX12-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB2_4
+; GFX12-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX12-NEXT: s_mov_b32 exec_lo, s2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX12-NEXT: v_mov_b32_e32 v6, v4
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB2_3
+; GFX12-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f32__offset__waterfall:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v8, 0x400, v4
+; GFX940-NEXT: s_mov_b64 s[2:3], exec
+; GFX940-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: buffer_load_dword v7, v4, s[4:7], 0 offen offset:1024
+; GFX940-NEXT: ; implicit-def: $vgpr4
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB2_1
+; GFX940-NEXT: ; %bb.2:
+; GFX940-NEXT: s_mov_b64 exec, s[2:3]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_max_f32_e32 v9, v5, v5
+; GFX940-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Loop Header: Depth=1
+; GFX940-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f32_e32 v4, v7, v7
+; GFX940-NEXT: v_min_f32_e32 v6, v4, v9
+; GFX940-NEXT: s_mov_b64 s[8:9], exec
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[6:7]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX940-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[4:7], 0 offen sc0
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB2_4
+; GFX940-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX940-NEXT: s_mov_b64 exec, s[8:9]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB2_3
+; GFX940-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f32__offset__waterfall:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: buffer_load_b32 v6, v4, s[4:7], 0 offen offset:1024
+; GFX11-NEXT: ; implicit-def: $vgpr4
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB2_1
+; GFX11-NEXT: ; %bb.2:
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: v_max_f32_e32 v8, v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Loop Header: Depth=1
+; GFX11-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f32_e32 v4, v6, v6
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f32_e32 v5, v4, v8
+; GFX11-NEXT: v_mov_b32_e32 v4, v5
+; GFX11-NEXT: v_mov_b32_e32 v5, v6
+; GFX11-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX11-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], 0 offen glc
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB2_4
+; GFX11-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX11-NEXT: v_mov_b32_e32 v6, v4
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB2_3
+; GFX11-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f32__offset__waterfall:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX10-NEXT: ; implicit-def: $vgpr4
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB2_1
+; GFX10-NEXT: ; %bb.2:
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: v_max_f32_e32 v8, v5, v5
+; GFX10-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Loop Header: Depth=1
+; GFX10-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f32_e32 v4, v6, v6
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_min_f32_e32 v5, v4, v8
+; GFX10-NEXT: v_mov_b32_e32 v4, v5
+; GFX10-NEXT: v_mov_b32_e32 v5, v6
+; GFX10-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX10-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB2_4
+; GFX10-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB2_3
+; GFX10-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f32__offset__waterfall:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v8, 0x400, v4
+; GFX90A-NEXT: s_mov_b64 s[6:7], exec
+; GFX90A-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_nop 0
+; GFX90A-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen offset:1024
+; GFX90A-NEXT: ; implicit-def: $vgpr4
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB2_1
+; GFX90A-NEXT: ; %bb.2:
+; GFX90A-NEXT: s_mov_b64 exec, s[6:7]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_max_f32_e32 v9, v5, v5
+; GFX90A-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Loop Header: Depth=1
+; GFX90A-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f32_e32 v4, v7, v7
+; GFX90A-NEXT: v_min_f32_e32 v6, v4, v9
+; GFX90A-NEXT: s_mov_b64 s[12:13], exec
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[6:7] op_sel:[0,1]
+; GFX90A-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX90A-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[8:11], 0 offen glc
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB2_4
+; GFX90A-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX90A-NEXT: s_mov_b64 exec, s[12:13]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB2_3
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f32__offset__waterfall:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v7, 0x400, v4
+; GFX908-NEXT: s_mov_b64 s[6:7], exec
+; GFX908-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_nop 0
+; GFX908-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX908-NEXT: ; implicit-def: $vgpr4
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB2_1
+; GFX908-NEXT: ; %bb.2:
+; GFX908-NEXT: s_mov_b64 exec, s[6:7]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_max_f32_e32 v8, v5, v5
+; GFX908-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Loop Header: Depth=1
+; GFX908-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f32_e32 v4, v6, v6
+; GFX908-NEXT: v_min_f32_e32 v5, v4, v8
+; GFX908-NEXT: v_mov_b32_e32 v4, v5
+; GFX908-NEXT: s_mov_b64 s[12:13], exec
+; GFX908-NEXT: v_mov_b32_e32 v5, v6
+; GFX908-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX908-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB2_4
+; GFX908-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX908-NEXT: s_mov_b64 exec, s[12:13]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB2_3
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f32__offset__waterfall:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x400, v4
+; GFX8-NEXT: s_mov_b64 s[6:7], exec
+; GFX8-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_nop 0
+; GFX8-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX8-NEXT: ; implicit-def: $vgpr4
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB2_1
+; GFX8-NEXT: ; %bb.2:
+; GFX8-NEXT: s_mov_b64 exec, s[6:7]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_mul_f32_e32 v8, 1.0, v5
+; GFX8-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Loop Header: Depth=1
+; GFX8-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mul_f32_e32 v4, 1.0, v6
+; GFX8-NEXT: v_min_f32_e32 v5, v4, v8
+; GFX8-NEXT: v_mov_b32_e32 v4, v5
+; GFX8-NEXT: s_mov_b64 s[12:13], exec
+; GFX8-NEXT: v_mov_b32_e32 v5, v6
+; GFX8-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX8-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB2_4
+; GFX8-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX8-NEXT: s_mov_b64 exec, s[12:13]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB2_3
+; GFX8-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f32__offset__waterfall:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v7, vcc, 0x400, v4
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX7-NEXT: ; implicit-def: $vgpr4
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_1
+; GFX7-NEXT: ; %bb.2:
+; GFX7-NEXT: s_mov_b64 exec, s[6:7]
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v5
+; GFX7-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Loop Header: Depth=1
+; GFX7-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v6
+; GFX7-NEXT: v_min_f32_e32 v5, v4, v8
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_mov_b64 s[12:13], exec
+; GFX7-NEXT: v_mov_b32_e32 v5, v6
+; GFX7-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX7-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_4
+; GFX7-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX7-NEXT: s_mov_b64 exec, s[12:13]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB2_3
+; GFX7-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v0, v4
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f32__offset__waterfall:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v7, vcc, 0x400, v4
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX6-NEXT: ; implicit-def: $vgpr4
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB2_1
+; GFX6-NEXT: ; %bb.2:
+; GFX6-NEXT: s_mov_b64 exec, s[6:7]
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: v_mul_f32_e32 v8, 1.0, v5
+; GFX6-NEXT: .LBB2_3: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Loop Header: Depth=1
+; GFX6-NEXT: ; Child Loop BB2_4 Depth 2
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v6
+; GFX6-NEXT: v_min_f32_e32 v5, v4, v8
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_mov_b64 s[12:13], exec
+; GFX6-NEXT: v_mov_b32_e32 v5, v6
+; GFX6-NEXT: .LBB2_4: ; Parent Loop BB2_3 Depth=1
+; GFX6-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB2_4
+; GFX6-NEXT: ; %bb.5: ; in Loop: Header=BB2_3 Depth=1
+; GFX6-NEXT: s_mov_b64 exec, s[12:13]
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB2_3
+; GFX6-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v0, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmin ptr addrspace(7) %gep, float %val syncscope("agent") seq_cst
+ ret float %result
+}
+
+; --------------------------------------------------------------------
+; double
+; --------------------------------------------------------------------
+
+define double @buffer_fat_ptr_agent_atomic_fmin_ret_f64__offset(ptr addrspace(7) inreg %ptr, double %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f64__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX12-NEXT: v_mov_b32_e32 v0, s4
+; GFX12-NEXT: s_addk_co_i32 s4, 0x800
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_mov_b32_e32 v6, s4
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[2:3], v[2:3]
+; GFX12-NEXT: buffer_load_b64 v[0:1], v0, s[0:3], null offen offset:2048
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v10, v1 :: v_dual_mov_b32 v9, v0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f64_e32 v[0:1], v[9:10], v[9:10]
+; GFX12-NEXT: v_min_num_f64_e32 v[7:8], v[0:1], v[4:5]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v0, v7 :: v_dual_mov_b32 v1, v8
+; GFX12-NEXT: v_dual_mov_b32 v2, v9 :: v_dual_mov_b32 v3, v10
+; GFX12-NEXT: buffer_atomic_cmpswap_b64 v[0:3], v6, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[9:10]
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB3_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f64__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, v0
+; GFX940-NEXT: v_mov_b32_e32 v0, s4
+; GFX940-NEXT: v_mov_b32_e32 v3, v1
+; GFX940-NEXT: buffer_load_dwordx2 v[0:1], v0, s[0:3], 0 offen offset:2048
+; GFX940-NEXT: s_add_i32 s6, s4, 0x800
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v6, s6
+; GFX940-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b64_e32 v[10:11], v[0:1]
+; GFX940-NEXT: v_max_f64 v[0:1], v[10:11], v[10:11]
+; GFX940-NEXT: v_min_f64 v[8:9], v[0:1], v[4:5]
+; GFX940-NEXT: v_mov_b64_e32 v[0:1], v[8:9]
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], v[10:11]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v6, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB3_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f64__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX11-NEXT: v_mov_b32_e32 v0, s4
+; GFX11-NEXT: s_addk_i32 s4, 0x800
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_mov_b32_e32 v6, s4
+; GFX11-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX11-NEXT: buffer_load_b64 v[0:1], v0, s[0:3], 0 offen offset:2048
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v10, v1 :: v_dual_mov_b32 v9, v0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f64 v[0:1], v[9:10], v[9:10]
+; GFX11-NEXT: v_min_f64 v[7:8], v[0:1], v[4:5]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v0, v7 :: v_dual_mov_b32 v1, v8
+; GFX11-NEXT: v_dual_mov_b32 v2, v9 :: v_dual_mov_b32 v3, v10
+; GFX11-NEXT: buffer_atomic_cmpswap_b64 v[0:3], v6, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[9:10]
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB3_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f64__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, v0
+; GFX10-NEXT: v_mov_b32_e32 v0, s8
+; GFX10-NEXT: v_mov_b32_e32 v3, v1
+; GFX10-NEXT: s_addk_i32 s8, 0x800
+; GFX10-NEXT: v_mov_b32_e32 v6, s8
+; GFX10-NEXT: buffer_load_dwordx2 v[0:1], v0, s[4:7], 0 offen offset:2048
+; GFX10-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v10, v1
+; GFX10-NEXT: v_mov_b32_e32 v9, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_max_f64 v[0:1], v[9:10], v[9:10]
+; GFX10-NEXT: v_min_f64 v[7:8], v[0:1], v[4:5]
+; GFX10-NEXT: v_mov_b32_e32 v0, v7
+; GFX10-NEXT: v_mov_b32_e32 v1, v8
+; GFX10-NEXT: v_mov_b32_e32 v2, v9
+; GFX10-NEXT: v_mov_b32_e32 v3, v10
+; GFX10-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v6, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[9:10]
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB3_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f64__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, v0
+; GFX90A-NEXT: v_mov_b32_e32 v0, s8
+; GFX90A-NEXT: v_mov_b32_e32 v3, v1
+; GFX90A-NEXT: buffer_load_dwordx2 v[0:1], v0, s[4:7], 0 offen offset:2048
+; GFX90A-NEXT: s_add_i32 s10, s8, 0x800
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX90A-NEXT: v_mov_b32_e32 v6, s10
+; GFX90A-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_mov_b32 v[10:11], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: v_max_f64 v[0:1], v[10:11], v[10:11]
+; GFX90A-NEXT: v_min_f64 v[8:9], v[0:1], v[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[8:9], v[8:9] op_sel:[0,1]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[10:11], v[10:11] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v6, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[10:11]
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB3_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f64__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, v0
+; GFX908-NEXT: v_mov_b32_e32 v0, s8
+; GFX908-NEXT: v_mov_b32_e32 v3, v1
+; GFX908-NEXT: buffer_load_dwordx2 v[0:1], v0, s[4:7], 0 offen offset:2048
+; GFX908-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX908-NEXT: s_add_i32 s10, s8, 0x800
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_mov_b32_e32 v6, s10
+; GFX908-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v10, v1
+; GFX908-NEXT: v_mov_b32_e32 v9, v0
+; GFX908-NEXT: v_max_f64 v[0:1], v[9:10], v[9:10]
+; GFX908-NEXT: v_min_f64 v[7:8], v[0:1], v[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v7
+; GFX908-NEXT: v_mov_b32_e32 v1, v8
+; GFX908-NEXT: v_mov_b32_e32 v2, v9
+; GFX908-NEXT: v_mov_b32_e32 v3, v10
+; GFX908-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v6, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[9:10]
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB3_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f64__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, v0
+; GFX8-NEXT: v_mov_b32_e32 v0, s8
+; GFX8-NEXT: v_mov_b32_e32 v3, v1
+; GFX8-NEXT: buffer_load_dwordx2 v[0:1], v0, s[4:7], 0 offen offset:2048
+; GFX8-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX8-NEXT: s_add_i32 s10, s8, 0x800
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_mov_b32_e32 v6, s10
+; GFX8-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v10, v1
+; GFX8-NEXT: v_mov_b32_e32 v9, v0
+; GFX8-NEXT: v_max_f64 v[0:1], v[9:10], v[9:10]
+; GFX8-NEXT: v_min_f64 v[7:8], v[0:1], v[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v7
+; GFX8-NEXT: v_mov_b32_e32 v1, v8
+; GFX8-NEXT: v_mov_b32_e32 v2, v9
+; GFX8-NEXT: v_mov_b32_e32 v3, v10
+; GFX8-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v6, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[9:10]
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB3_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f64__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, v0
+; GFX7-NEXT: v_mov_b32_e32 v0, s8
+; GFX7-NEXT: v_mov_b32_e32 v3, v1
+; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v0, s[4:7], 0 offen offset:2048
+; GFX7-NEXT: s_add_i32 s10, s8, 0x800
+; GFX7-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mov_b32_e32 v6, s10
+; GFX7-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v10, v1
+; GFX7-NEXT: v_mov_b32_e32 v9, v0
+; GFX7-NEXT: v_max_f64 v[0:1], v[9:10], v[9:10]
+; GFX7-NEXT: v_min_f64 v[7:8], v[0:1], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v7
+; GFX7-NEXT: v_mov_b32_e32 v1, v8
+; GFX7-NEXT: v_mov_b32_e32 v2, v9
+; GFX7-NEXT: v_mov_b32_e32 v3, v10
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v6, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[9:10]
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB3_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f64__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, v0
+; GFX6-NEXT: v_mov_b32_e32 v0, s8
+; GFX6-NEXT: v_mov_b32_e32 v3, v1
+; GFX6-NEXT: buffer_load_dwordx2 v[0:1], v0, s[4:7], 0 offen offset:2048
+; GFX6-NEXT: s_add_i32 s10, s8, 0x800
+; GFX6-NEXT: v_max_f64 v[4:5], v[2:3], v[2:3]
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mov_b32_e32 v6, s10
+; GFX6-NEXT: .LBB3_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v10, v1
+; GFX6-NEXT: v_mov_b32_e32 v9, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_max_f64 v[0:1], v[9:10], v[9:10]
+; GFX6-NEXT: v_min_f64 v[7:8], v[0:1], v[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v0, v7
+; GFX6-NEXT: v_mov_b32_e32 v1, v8
+; GFX6-NEXT: v_mov_b32_e32 v2, v9
+; GFX6-NEXT: v_mov_b32_e32 v3, v10
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v6, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[9:10]
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB3_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmin ptr addrspace(7) %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+define void @buffer_fat_ptr_agent_atomic_fmin_noret_f64__offset(ptr addrspace(7) inreg %ptr, double %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f64__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v2, s4
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[0:1], v[0:1]
+; GFX12-NEXT: s_addk_co_i32 s4, 0x800
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_mov_b32_e32 v6, s4
+; GFX12-NEXT: buffer_load_b64 v[2:3], v2, s[0:3], null offen offset:2048
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f64_e32 v[0:1], v[2:3], v[2:3]
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_min_num_f64_e32 v[0:1], v[0:1], v[4:5]
+; GFX12-NEXT: v_dual_mov_b32 v10, v3 :: v_dual_mov_b32 v9, v2
+; GFX12-NEXT: v_dual_mov_b32 v8, v1 :: v_dual_mov_b32 v7, v0
+; GFX12-NEXT: buffer_atomic_cmpswap_b64 v[7:10], v6, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[7:8], v[2:3]
+; GFX12-NEXT: v_dual_mov_b32 v2, v7 :: v_dual_mov_b32 v3, v8
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB4_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f64__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v2, s4
+; GFX940-NEXT: buffer_load_dwordx2 v[2:3], v2, s[0:3], 0 offen offset:2048
+; GFX940-NEXT: s_add_i32 s6, s4, 0x800
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
+; GFX940-NEXT: v_mov_b32_e32 v6, s6
+; GFX940-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX940-NEXT: v_min_f64 v[0:1], v[0:1], v[4:5]
+; GFX940-NEXT: v_mov_b64_e32 v[10:11], v[2:3]
+; GFX940-NEXT: v_mov_b64_e32 v[8:9], v[0:1]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v6, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[2:3]
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], v[8:9]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB4_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f64__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v2, s4
+; GFX11-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
+; GFX11-NEXT: s_addk_i32 s4, 0x800
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v6, s4
+; GFX11-NEXT: buffer_load_b64 v[2:3], v2, s[0:3], 0 offen offset:2048
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_min_f64 v[0:1], v[0:1], v[4:5]
+; GFX11-NEXT: v_dual_mov_b32 v10, v3 :: v_dual_mov_b32 v9, v2
+; GFX11-NEXT: v_dual_mov_b32 v8, v1 :: v_dual_mov_b32 v7, v0
+; GFX11-NEXT: buffer_atomic_cmpswap_b64 v[7:10], v6, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[7:8], v[2:3]
+; GFX11-NEXT: v_dual_mov_b32 v2, v7 :: v_dual_mov_b32 v3, v8
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB4_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f64__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v2, s8
+; GFX10-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
+; GFX10-NEXT: s_addk_i32 s8, 0x800
+; GFX10-NEXT: v_mov_b32_e32 v6, s8
+; GFX10-NEXT: buffer_load_dwordx2 v[2:3], v2, s[4:7], 0 offen offset:2048
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_min_f64 v[0:1], v[0:1], v[4:5]
+; GFX10-NEXT: v_mov_b32_e32 v10, v3
+; GFX10-NEXT: v_mov_b32_e32 v9, v2
+; GFX10-NEXT: v_mov_b32_e32 v8, v1
+; GFX10-NEXT: v_mov_b32_e32 v7, v0
+; GFX10-NEXT: buffer_atomic_cmpswap_x2 v[7:10], v6, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[7:8], v[2:3]
+; GFX10-NEXT: v_mov_b32_e32 v2, v7
+; GFX10-NEXT: v_mov_b32_e32 v3, v8
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB4_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f64__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v2, s8
+; GFX90A-NEXT: buffer_load_dwordx2 v[2:3], v2, s[4:7], 0 offen offset:2048
+; GFX90A-NEXT: s_add_i32 s10, s8, 0x800
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
+; GFX90A-NEXT: v_mov_b32_e32 v6, s10
+; GFX90A-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX90A-NEXT: v_min_f64 v[0:1], v[0:1], v[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[10:11], v[2:3], v[2:3] op_sel:[0,1]
+; GFX90A-NEXT: v_pk_mov_b32 v[8:9], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap_x2 v[8:11], v6, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[8:9], v[2:3]
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[8:9], v[8:9] op_sel:[0,1]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB4_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f64__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v2, s8
+; GFX908-NEXT: buffer_load_dwordx2 v[2:3], v2, s[4:7], 0 offen offset:2048
+; GFX908-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
+; GFX908-NEXT: s_add_i32 s10, s8, 0x800
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_mov_b32_e32 v6, s10
+; GFX908-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX908-NEXT: v_min_f64 v[0:1], v[0:1], v[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v10, v3
+; GFX908-NEXT: v_mov_b32_e32 v9, v2
+; GFX908-NEXT: v_mov_b32_e32 v8, v1
+; GFX908-NEXT: v_mov_b32_e32 v7, v0
+; GFX908-NEXT: buffer_atomic_cmpswap_x2 v[7:10], v6, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[7:8], v[2:3]
+; GFX908-NEXT: v_mov_b32_e32 v2, v7
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: v_mov_b32_e32 v3, v8
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB4_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f64__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v2, s8
+; GFX8-NEXT: buffer_load_dwordx2 v[2:3], v2, s[4:7], 0 offen offset:2048
+; GFX8-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
+; GFX8-NEXT: s_add_i32 s10, s8, 0x800
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_mov_b32_e32 v6, s10
+; GFX8-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX8-NEXT: v_min_f64 v[0:1], v[0:1], v[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v10, v3
+; GFX8-NEXT: v_mov_b32_e32 v9, v2
+; GFX8-NEXT: v_mov_b32_e32 v8, v1
+; GFX8-NEXT: v_mov_b32_e32 v7, v0
+; GFX8-NEXT: buffer_atomic_cmpswap_x2 v[7:10], v6, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[7:8], v[2:3]
+; GFX8-NEXT: v_mov_b32_e32 v2, v7
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v3, v8
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB4_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f64__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dwordx2 v[2:3], v2, s[4:7], 0 offen offset:2048
+; GFX7-NEXT: s_add_i32 s10, s8, 0x800
+; GFX7-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mov_b32_e32 v6, s10
+; GFX7-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX7-NEXT: v_min_f64 v[0:1], v[0:1], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v10, v3
+; GFX7-NEXT: v_mov_b32_e32 v9, v2
+; GFX7-NEXT: v_mov_b32_e32 v8, v1
+; GFX7-NEXT: v_mov_b32_e32 v7, v0
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[7:10], v6, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[7:8], v[2:3]
+; GFX7-NEXT: v_mov_b32_e32 v2, v7
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v3, v8
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB4_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f64__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dwordx2 v[2:3], v2, s[4:7], 0 offen offset:2048
+; GFX6-NEXT: s_add_i32 s10, s8, 0x800
+; GFX6-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mov_b32_e32 v6, s10
+; GFX6-NEXT: .LBB4_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX6-NEXT: v_min_f64 v[0:1], v[0:1], v[4:5]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v10, v3
+; GFX6-NEXT: v_mov_b32_e32 v9, v2
+; GFX6-NEXT: v_mov_b32_e32 v8, v1
+; GFX6-NEXT: v_mov_b32_e32 v7, v0
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[7:10], v6, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[7:8], v[2:3]
+; GFX6-NEXT: v_mov_b32_e32 v2, v7
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v3, v8
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB4_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(7) %ptr, i32 256
+ %unused = atomicrmw fmin ptr addrspace(7) %gep, double %val syncscope("agent") seq_cst
+ ret void
+}
+
+define double @buffer_fat_ptr_agent_atomic_fmin_ret_f64__offset__waterfall(ptr addrspace(7) %ptr, double %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f64__offset__waterfall:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v8, v3 :: v_dual_mov_b32 v7, v2
+; GFX12-NEXT: v_dual_mov_b32 v10, v1 :: v_dual_mov_b32 v9, v0
+; GFX12-NEXT: v_add_nc_u32_e32 v15, 0x800, v4
+; GFX12-NEXT: s_mov_b32 s1, exec_lo
+; GFX12-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_readfirstlane_b32 s4, v9
+; GFX12-NEXT: v_readfirstlane_b32 s5, v10
+; GFX12-NEXT: v_readfirstlane_b32 s6, v7
+; GFX12-NEXT: v_readfirstlane_b32 s7, v8
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[9:10]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[7:8]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: buffer_load_b64 v[13:14], v4, s[4:7], null offen offset:2048
+; GFX12-NEXT: ; implicit-def: $vgpr4
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB5_1
+; GFX12-NEXT: ; %bb.2:
+; GFX12-NEXT: s_mov_b32 exec_lo, s1
+; GFX12-NEXT: v_max_num_f64_e32 v[4:5], v[5:6], v[5:6]
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Loop Header: Depth=1
+; GFX12-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_max_num_f64_e32 v[0:1], v[13:14], v[13:14]
+; GFX12-NEXT: s_mov_b32 s2, exec_lo
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_min_num_f64_e32 v[11:12], v[0:1], v[4:5]
+; GFX12-NEXT: v_dual_mov_b32 v0, v11 :: v_dual_mov_b32 v1, v12
+; GFX12-NEXT: v_dual_mov_b32 v2, v13 :: v_dual_mov_b32 v3, v14
+; GFX12-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX12-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX12-NEXT: v_readfirstlane_b32 s4, v9
+; GFX12-NEXT: v_readfirstlane_b32 s5, v10
+; GFX12-NEXT: v_readfirstlane_b32 s6, v7
+; GFX12-NEXT: v_readfirstlane_b32 s7, v8
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[9:10]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[7:8]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: buffer_atomic_cmpswap_b64 v[0:3], v15, s[4:7], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB5_4
+; GFX12-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX12-NEXT: s_mov_b32 exec_lo, s2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[13:14]
+; GFX12-NEXT: v_dual_mov_b32 v14, v1 :: v_dual_mov_b32 v13, v0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB5_3
+; GFX12-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f64__offset__waterfall:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v11, v6
+; GFX940-NEXT: v_mov_b32_e32 v10, v5
+; GFX940-NEXT: v_mov_b32_e32 v9, v3
+; GFX940-NEXT: v_mov_b32_e32 v8, v2
+; GFX940-NEXT: v_mov_b32_e32 v7, v1
+; GFX940-NEXT: v_mov_b32_e32 v6, v0
+; GFX940-NEXT: v_add_u32_e32 v14, 0x800, v4
+; GFX940-NEXT: s_mov_b64 s[2:3], exec
+; GFX940-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: v_readfirstlane_b32 s4, v6
+; GFX940-NEXT: v_readfirstlane_b32 s5, v7
+; GFX940-NEXT: v_readfirstlane_b32 s6, v8
+; GFX940-NEXT: v_readfirstlane_b32 s7, v9
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[6:7]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[8:9]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: buffer_load_dwordx2 v[12:13], v4, s[4:7], 0 offen offset:2048
+; GFX940-NEXT: ; implicit-def: $vgpr4
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB5_1
+; GFX940-NEXT: ; %bb.2:
+; GFX940-NEXT: s_mov_b64 exec, s[2:3]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_max_f64 v[4:5], v[10:11], v[10:11]
+; GFX940-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Loop Header: Depth=1
+; GFX940-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f64 v[0:1], v[12:13], v[12:13]
+; GFX940-NEXT: v_min_f64 v[10:11], v[0:1], v[4:5]
+; GFX940-NEXT: v_mov_b64_e32 v[0:1], v[10:11]
+; GFX940-NEXT: s_mov_b64 s[8:9], exec
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], v[12:13]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX940-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX940-NEXT: v_readfirstlane_b32 s4, v6
+; GFX940-NEXT: v_readfirstlane_b32 s5, v7
+; GFX940-NEXT: v_readfirstlane_b32 s6, v8
+; GFX940-NEXT: v_readfirstlane_b32 s7, v9
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[6:7]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[8:9]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v14, s[4:7], 0 offen sc0
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB5_4
+; GFX940-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX940-NEXT: s_mov_b64 exec, s[8:9]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[12:13]
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b64_e32 v[12:13], v[0:1]
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB5_3
+; GFX940-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f64__offset__waterfall:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v8, v3 :: v_dual_mov_b32 v7, v2
+; GFX11-NEXT: v_dual_mov_b32 v10, v1 :: v_dual_mov_b32 v9, v0
+; GFX11-NEXT: v_add_nc_u32_e32 v15, 0x800, v4
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_readfirstlane_b32 s4, v9
+; GFX11-NEXT: v_readfirstlane_b32 s5, v10
+; GFX11-NEXT: v_readfirstlane_b32 s6, v7
+; GFX11-NEXT: v_readfirstlane_b32 s7, v8
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[9:10]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[7:8]
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: buffer_load_b64 v[13:14], v4, s[4:7], 0 offen offset:2048
+; GFX11-NEXT: ; implicit-def: $vgpr4
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB5_1
+; GFX11-NEXT: ; %bb.2:
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: v_max_f64 v[4:5], v[5:6], v[5:6]
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Loop Header: Depth=1
+; GFX11-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_max_f64 v[0:1], v[13:14], v[13:14]
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_min_f64 v[11:12], v[0:1], v[4:5]
+; GFX11-NEXT: v_dual_mov_b32 v0, v11 :: v_dual_mov_b32 v1, v12
+; GFX11-NEXT: v_dual_mov_b32 v2, v13 :: v_dual_mov_b32 v3, v14
+; GFX11-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX11-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX11-NEXT: v_readfirstlane_b32 s4, v9
+; GFX11-NEXT: v_readfirstlane_b32 s5, v10
+; GFX11-NEXT: v_readfirstlane_b32 s6, v7
+; GFX11-NEXT: v_readfirstlane_b32 s7, v8
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[9:10]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[7:8]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_atomic_cmpswap_b64 v[0:3], v15, s[4:7], 0 offen glc
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB5_4
+; GFX11-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[13:14]
+; GFX11-NEXT: v_dual_mov_b32 v14, v1 :: v_dual_mov_b32 v13, v0
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB5_3
+; GFX11-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f64__offset__waterfall:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v8, v3
+; GFX10-NEXT: v_mov_b32_e32 v7, v2
+; GFX10-NEXT: v_mov_b32_e32 v10, v1
+; GFX10-NEXT: v_mov_b32_e32 v9, v0
+; GFX10-NEXT: v_add_nc_u32_e32 v15, 0x800, v4
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_readfirstlane_b32 s8, v9
+; GFX10-NEXT: v_readfirstlane_b32 s9, v10
+; GFX10-NEXT: v_readfirstlane_b32 s10, v7
+; GFX10-NEXT: v_readfirstlane_b32 s11, v8
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[9:10]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[7:8]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: buffer_load_dwordx2 v[13:14], v4, s[8:11], 0 offen offset:2048
+; GFX10-NEXT: ; implicit-def: $vgpr4
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB5_1
+; GFX10-NEXT: ; %bb.2:
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: v_max_f64 v[4:5], v[5:6], v[5:6]
+; GFX10-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Loop Header: Depth=1
+; GFX10-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f64 v[0:1], v[13:14], v[13:14]
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_min_f64 v[11:12], v[0:1], v[4:5]
+; GFX10-NEXT: v_mov_b32_e32 v0, v11
+; GFX10-NEXT: v_mov_b32_e32 v1, v12
+; GFX10-NEXT: v_mov_b32_e32 v2, v13
+; GFX10-NEXT: v_mov_b32_e32 v3, v14
+; GFX10-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX10-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX10-NEXT: v_readfirstlane_b32 s8, v9
+; GFX10-NEXT: v_readfirstlane_b32 s9, v10
+; GFX10-NEXT: v_readfirstlane_b32 s10, v7
+; GFX10-NEXT: v_readfirstlane_b32 s11, v8
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[9:10]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[7:8]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v15, s[8:11], 0 offen glc
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB5_4
+; GFX10-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[13:14]
+; GFX10-NEXT: v_mov_b32_e32 v14, v1
+; GFX10-NEXT: v_mov_b32_e32 v13, v0
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB5_3
+; GFX10-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f64__offset__waterfall:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v11, v6
+; GFX90A-NEXT: v_mov_b32_e32 v10, v5
+; GFX90A-NEXT: v_mov_b32_e32 v9, v3
+; GFX90A-NEXT: v_mov_b32_e32 v8, v2
+; GFX90A-NEXT: v_mov_b32_e32 v7, v1
+; GFX90A-NEXT: v_mov_b32_e32 v6, v0
+; GFX90A-NEXT: v_add_u32_e32 v14, 0x800, v4
+; GFX90A-NEXT: s_mov_b64 s[6:7], exec
+; GFX90A-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v6
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v7
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v8
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v9
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[6:7]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[8:9]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_nop 0
+; GFX90A-NEXT: buffer_load_dwordx2 v[12:13], v4, s[8:11], 0 offen offset:2048
+; GFX90A-NEXT: ; implicit-def: $vgpr4
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB5_1
+; GFX90A-NEXT: ; %bb.2:
+; GFX90A-NEXT: s_mov_b64 exec, s[6:7]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_max_f64 v[4:5], v[10:11], v[10:11]
+; GFX90A-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Loop Header: Depth=1
+; GFX90A-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f64 v[0:1], v[12:13], v[12:13]
+; GFX90A-NEXT: v_min_f64 v[10:11], v[0:1], v[4:5]
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[10:11], v[10:11] op_sel:[0,1]
+; GFX90A-NEXT: s_mov_b64 s[12:13], exec
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[12:13], v[12:13] op_sel:[0,1]
+; GFX90A-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX90A-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v6
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v7
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v8
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v9
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[6:7]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[8:9]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v14, s[8:11], 0 offen glc
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB5_4
+; GFX90A-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX90A-NEXT: s_mov_b64 exec, s[12:13]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[12:13]
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_pk_mov_b32 v[12:13], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB5_3
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f64__offset__waterfall:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v8, v3
+; GFX908-NEXT: v_mov_b32_e32 v7, v2
+; GFX908-NEXT: v_mov_b32_e32 v10, v1
+; GFX908-NEXT: v_mov_b32_e32 v9, v0
+; GFX908-NEXT: v_add_u32_e32 v15, 0x800, v4
+; GFX908-NEXT: s_mov_b64 s[6:7], exec
+; GFX908-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: v_readfirstlane_b32 s8, v9
+; GFX908-NEXT: v_readfirstlane_b32 s9, v10
+; GFX908-NEXT: v_readfirstlane_b32 s10, v7
+; GFX908-NEXT: v_readfirstlane_b32 s11, v8
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_nop 0
+; GFX908-NEXT: buffer_load_dwordx2 v[13:14], v4, s[8:11], 0 offen offset:2048
+; GFX908-NEXT: ; implicit-def: $vgpr4
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_1
+; GFX908-NEXT: ; %bb.2:
+; GFX908-NEXT: s_mov_b64 exec, s[6:7]
+; GFX908-NEXT: v_max_f64 v[4:5], v[5:6], v[5:6]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Loop Header: Depth=1
+; GFX908-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f64 v[0:1], v[13:14], v[13:14]
+; GFX908-NEXT: s_mov_b64 s[12:13], exec
+; GFX908-NEXT: v_min_f64 v[11:12], v[0:1], v[4:5]
+; GFX908-NEXT: v_mov_b32_e32 v0, v11
+; GFX908-NEXT: v_mov_b32_e32 v1, v12
+; GFX908-NEXT: v_mov_b32_e32 v2, v13
+; GFX908-NEXT: v_mov_b32_e32 v3, v14
+; GFX908-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX908-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX908-NEXT: v_readfirstlane_b32 s8, v9
+; GFX908-NEXT: v_readfirstlane_b32 s9, v10
+; GFX908-NEXT: v_readfirstlane_b32 s10, v7
+; GFX908-NEXT: v_readfirstlane_b32 s11, v8
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v15, s[8:11], 0 offen glc
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_4
+; GFX908-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX908-NEXT: s_mov_b64 exec, s[12:13]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[13:14]
+; GFX908-NEXT: v_mov_b32_e32 v14, v1
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v13, v0
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB5_3
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f64__offset__waterfall:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v8, v3
+; GFX8-NEXT: v_mov_b32_e32 v7, v2
+; GFX8-NEXT: v_mov_b32_e32 v10, v1
+; GFX8-NEXT: v_mov_b32_e32 v9, v0
+; GFX8-NEXT: v_add_u32_e32 v15, vcc, 0x800, v4
+; GFX8-NEXT: s_mov_b64 s[6:7], exec
+; GFX8-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: v_readfirstlane_b32 s8, v9
+; GFX8-NEXT: v_readfirstlane_b32 s9, v10
+; GFX8-NEXT: v_readfirstlane_b32 s10, v7
+; GFX8-NEXT: v_readfirstlane_b32 s11, v8
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_nop 0
+; GFX8-NEXT: buffer_load_dwordx2 v[13:14], v4, s[8:11], 0 offen offset:2048
+; GFX8-NEXT: ; implicit-def: $vgpr4
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_1
+; GFX8-NEXT: ; %bb.2:
+; GFX8-NEXT: s_mov_b64 exec, s[6:7]
+; GFX8-NEXT: v_max_f64 v[4:5], v[5:6], v[5:6]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Loop Header: Depth=1
+; GFX8-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f64 v[0:1], v[13:14], v[13:14]
+; GFX8-NEXT: s_mov_b64 s[12:13], exec
+; GFX8-NEXT: v_min_f64 v[11:12], v[0:1], v[4:5]
+; GFX8-NEXT: v_mov_b32_e32 v0, v11
+; GFX8-NEXT: v_mov_b32_e32 v1, v12
+; GFX8-NEXT: v_mov_b32_e32 v2, v13
+; GFX8-NEXT: v_mov_b32_e32 v3, v14
+; GFX8-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX8-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX8-NEXT: v_readfirstlane_b32 s8, v9
+; GFX8-NEXT: v_readfirstlane_b32 s9, v10
+; GFX8-NEXT: v_readfirstlane_b32 s10, v7
+; GFX8-NEXT: v_readfirstlane_b32 s11, v8
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v15, s[8:11], 0 offen glc
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_4
+; GFX8-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX8-NEXT: s_mov_b64 exec, s[12:13]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[13:14]
+; GFX8-NEXT: v_mov_b32_e32 v14, v1
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v13, v0
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB5_3
+; GFX8-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f64__offset__waterfall:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v8, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v2
+; GFX7-NEXT: v_mov_b32_e32 v10, v1
+; GFX7-NEXT: v_mov_b32_e32 v9, v0
+; GFX7-NEXT: v_add_i32_e32 v15, vcc, 0x800, v4
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_readfirstlane_b32 s8, v9
+; GFX7-NEXT: v_readfirstlane_b32 s9, v10
+; GFX7-NEXT: v_readfirstlane_b32 s10, v7
+; GFX7-NEXT: v_readfirstlane_b32 s11, v8
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: buffer_load_dwordx2 v[13:14], v4, s[8:11], 0 offen offset:2048
+; GFX7-NEXT: ; implicit-def: $vgpr4
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_1
+; GFX7-NEXT: ; %bb.2:
+; GFX7-NEXT: s_mov_b64 exec, s[6:7]
+; GFX7-NEXT: v_max_f64 v[4:5], v[5:6], v[5:6]
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Loop Header: Depth=1
+; GFX7-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_max_f64 v[0:1], v[13:14], v[13:14]
+; GFX7-NEXT: s_mov_b64 s[12:13], exec
+; GFX7-NEXT: v_min_f64 v[11:12], v[0:1], v[4:5]
+; GFX7-NEXT: v_mov_b32_e32 v0, v11
+; GFX7-NEXT: v_mov_b32_e32 v1, v12
+; GFX7-NEXT: v_mov_b32_e32 v2, v13
+; GFX7-NEXT: v_mov_b32_e32 v3, v14
+; GFX7-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX7-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX7-NEXT: v_readfirstlane_b32 s8, v9
+; GFX7-NEXT: v_readfirstlane_b32 s9, v10
+; GFX7-NEXT: v_readfirstlane_b32 s10, v7
+; GFX7-NEXT: v_readfirstlane_b32 s11, v8
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v15, s[8:11], 0 offen glc
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_4
+; GFX7-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX7-NEXT: s_mov_b64 exec, s[12:13]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[13:14]
+; GFX7-NEXT: v_mov_b32_e32 v14, v1
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v13, v0
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB5_3
+; GFX7-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f64__offset__waterfall:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v8, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v2
+; GFX6-NEXT: v_mov_b32_e32 v10, v1
+; GFX6-NEXT: v_mov_b32_e32 v9, v0
+; GFX6-NEXT: v_add_i32_e32 v15, vcc, 0x800, v4
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: .LBB5_1: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_readfirstlane_b32 s8, v9
+; GFX6-NEXT: v_readfirstlane_b32 s9, v10
+; GFX6-NEXT: v_readfirstlane_b32 s10, v7
+; GFX6-NEXT: v_readfirstlane_b32 s11, v8
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: buffer_load_dwordx2 v[13:14], v4, s[8:11], 0 offen offset:2048
+; GFX6-NEXT: ; implicit-def: $vgpr4
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB5_1
+; GFX6-NEXT: ; %bb.2:
+; GFX6-NEXT: s_mov_b64 exec, s[6:7]
+; GFX6-NEXT: v_max_f64 v[4:5], v[5:6], v[5:6]
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: .LBB5_3: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Loop Header: Depth=1
+; GFX6-NEXT: ; Child Loop BB5_4 Depth 2
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: v_max_f64 v[0:1], v[13:14], v[13:14]
+; GFX6-NEXT: s_mov_b64 s[12:13], exec
+; GFX6-NEXT: v_min_f64 v[11:12], v[0:1], v[4:5]
+; GFX6-NEXT: v_mov_b32_e32 v0, v11
+; GFX6-NEXT: v_mov_b32_e32 v1, v12
+; GFX6-NEXT: v_mov_b32_e32 v2, v13
+; GFX6-NEXT: v_mov_b32_e32 v3, v14
+; GFX6-NEXT: .LBB5_4: ; Parent Loop BB5_3 Depth=1
+; GFX6-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX6-NEXT: v_readfirstlane_b32 s8, v9
+; GFX6-NEXT: v_readfirstlane_b32 s9, v10
+; GFX6-NEXT: v_readfirstlane_b32 s10, v7
+; GFX6-NEXT: v_readfirstlane_b32 s11, v8
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[9:10]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[7:8]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: buffer_atomic_cmpswap_x2 v[0:3], v15, s[8:11], 0 offen glc
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB5_4
+; GFX6-NEXT: ; %bb.5: ; in Loop: Header=BB5_3 Depth=1
+; GFX6-NEXT: s_mov_b64 exec, s[12:13]
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[13:14]
+; GFX6-NEXT: v_mov_b32_e32 v14, v1
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v13, v0
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB5_3
+; GFX6-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr double, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmin ptr addrspace(7) %gep, double %val syncscope("agent") seq_cst
+ ret double %result
+}
+
+; --------------------------------------------------------------------
+; half
+; --------------------------------------------------------------------
+
+define half @buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset(ptr addrspace(7) inreg %ptr, half %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_addk_co_i32 s4, 0x200
+; GFX12-NEXT: v_max_num_f16_e32 v5, v0, v0
+; GFX12-NEXT: s_and_b32 s4, s4, -4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_mov_b32_e32 v4, s4
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: buffer_load_b32 v1, v4, s[0:3], null offen
+; GFX12-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v0, v0, v0
+; GFX12-NEXT: v_min_num_f16_e32 v0, v0, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX12-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[2:3], v4, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB6_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_addk_i32 s4, 0x200
+; GFX940-NEXT: s_and_b32 s4, s4, -4
+; GFX940-NEXT: v_mov_b32_e32 v4, s4
+; GFX940-NEXT: buffer_load_dword v1, v4, s[0:3], 0 offen
+; GFX940-NEXT: s_mov_b32 s6, 0xffffff
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_max_f16_e32 v5, v0, v0
+; GFX940-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX940-NEXT: v_min_f16_sdwa v0, v0, v5 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_and_or_b32 v0, v1, s6, v0
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], v[0:1]
+; GFX940-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB6_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_addk_i32 s4, 0x200
+; GFX11-NEXT: v_max_f16_e32 v5, v0, v0
+; GFX11-NEXT: s_and_b32 s4, s4, -4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v4, s4
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: buffer_load_b32 v1, v4, s[0:3], 0 offen
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v0, v0, v0
+; GFX11-NEXT: v_min_f16_e32 v0, v0, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX11-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[2:3], v4, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB6_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_addk_i32 s8, 0x200
+; GFX10-NEXT: v_max_f16_e32 v5, v0, v0
+; GFX10-NEXT: s_and_b32 s8, s8, -4
+; GFX10-NEXT: v_mov_b32_e32 v4, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX10-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_min_f16_e32 v0, v0, v5
+; GFX10-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX10-NEXT: v_mov_b32_e32 v3, v1
+; GFX10-NEXT: v_mov_b32_e32 v2, v0
+; GFX10-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB6_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: s_addk_i32 s8, 0x200
+; GFX90A-NEXT: s_and_b32 s8, s8, -4
+; GFX90A-NEXT: v_mov_b32_e32 v4, s8
+; GFX90A-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX90A-NEXT: s_mov_b32 s10, 0xffffff
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_max_f16_e32 v5, v0, v0
+; GFX90A-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX90A-NEXT: v_min_f16_sdwa v0, v0, v5 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_and_or_b32 v0, v1, s10, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB6_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: s_addk_i32 s8, 0x200
+; GFX908-NEXT: s_and_b32 s8, s8, -4
+; GFX908-NEXT: v_mov_b32_e32 v4, s8
+; GFX908-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX908-NEXT: s_mov_b32 s10, 0xffffff
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_max_f16_e32 v5, v0, v0
+; GFX908-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX908-NEXT: v_min_f16_sdwa v0, v0, v5 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_and_or_b32 v0, v1, s10, v0
+; GFX908-NEXT: v_mov_b32_e32 v3, v1
+; GFX908-NEXT: v_mov_b32_e32 v2, v0
+; GFX908-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB6_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_addk_i32 s8, 0x200
+; GFX8-NEXT: s_and_b32 s8, s8, -4
+; GFX8-NEXT: v_mov_b32_e32 v4, s8
+; GFX8-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_max_f16_e32 v5, v0, v0
+; GFX8-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffffff, v1
+; GFX8-NEXT: v_min_f16_sdwa v0, v0, v5 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_or_b32_e32 v0, v2, v0
+; GFX8-NEXT: v_mov_b32_e32 v3, v1
+; GFX8-NEXT: v_mov_b32_e32 v2, v0
+; GFX8-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB6_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_addk_i32 s8, 0x200
+; GFX7-NEXT: s_and_b32 s8, s8, -4
+; GFX7-NEXT: v_mov_b32_e32 v4, s8
+; GFX7-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v0
+; GFX7-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffffff, v1
+; GFX7-NEXT: v_min_f32_e32 v0, v0, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX7-NEXT: v_or_b32_e32 v0, v2, v0
+; GFX7-NEXT: v_mov_b32_e32 v3, v1
+; GFX7-NEXT: v_mov_b32_e32 v2, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB6_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_addk_i32 s8, 0x200
+; GFX6-NEXT: s_and_b32 s8, s8, -4
+; GFX6-NEXT: v_mov_b32_e32 v4, s8
+; GFX6-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v0
+; GFX6-NEXT: .LBB6_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffffff, v1
+; GFX6-NEXT: v_min_f32_e32 v0, v0, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX6-NEXT: v_or_b32_e32 v0, v2, v0
+; GFX6-NEXT: v_mov_b32_e32 v3, v1
+; GFX6-NEXT: v_mov_b32_e32 v2, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB6_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmin ptr addrspace(7) %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+define void @buffer_fat_ptr_agent_atomic_fmin_noret_f16__offset(ptr addrspace(7) inreg %ptr, half %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_addk_co_i32 s4, 0x200
+; GFX12-NEXT: v_max_num_f16_e32 v3, v0, v0
+; GFX12-NEXT: s_and_b32 s4, s4, -4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_mov_b32_e32 v2, s4
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: buffer_load_b32 v1, v2, s[0:3], null offen
+; GFX12-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v0, v0, v0
+; GFX12-NEXT: v_min_num_f16_e32 v0, v0, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX12-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v2, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v4
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB7_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_addk_i32 s4, 0x200
+; GFX940-NEXT: s_and_b32 s4, s4, -4
+; GFX940-NEXT: v_mov_b32_e32 v2, s4
+; GFX940-NEXT: buffer_load_dword v1, v2, s[0:3], 0 offen
+; GFX940-NEXT: s_mov_b32 s6, 0xffffff
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_max_f16_e32 v3, v0, v0
+; GFX940-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX940-NEXT: v_min_f16_sdwa v0, v0, v3 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_and_or_b32 v0, v1, s6, v0
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[0:1]
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB7_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_addk_i32 s4, 0x200
+; GFX11-NEXT: v_max_f16_e32 v3, v0, v0
+; GFX11-NEXT: s_and_b32 s4, s4, -4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v2, s4
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: buffer_load_b32 v1, v2, s[0:3], 0 offen
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v0, v0, v0
+; GFX11-NEXT: v_min_f16_e32 v0, v0, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX11-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v2, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v4
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB7_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_addk_i32 s8, 0x200
+; GFX10-NEXT: v_max_f16_e32 v3, v0, v0
+; GFX10-NEXT: s_and_b32 s8, s8, -4
+; GFX10-NEXT: v_mov_b32_e32 v2, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX10-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_min_f16_e32 v0, v0, v3
+; GFX10-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX10-NEXT: v_mov_b32_e32 v5, v1
+; GFX10-NEXT: v_mov_b32_e32 v4, v0
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v4
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB7_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: s_addk_i32 s8, 0x200
+; GFX90A-NEXT: s_and_b32 s8, s8, -4
+; GFX90A-NEXT: v_mov_b32_e32 v2, s8
+; GFX90A-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX90A-NEXT: s_mov_b32 s10, 0xffffff
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_max_f16_e32 v3, v0, v0
+; GFX90A-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX90A-NEXT: v_min_f16_sdwa v0, v0, v3 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_and_or_b32 v0, v1, s10, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB7_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: s_addk_i32 s8, 0x200
+; GFX908-NEXT: s_and_b32 s8, s8, -4
+; GFX908-NEXT: v_mov_b32_e32 v2, s8
+; GFX908-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX908-NEXT: s_mov_b32 s10, 0xffffff
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_max_f16_e32 v3, v0, v0
+; GFX908-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX908-NEXT: v_min_f16_sdwa v0, v0, v3 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_and_or_b32 v0, v1, s10, v0
+; GFX908-NEXT: v_mov_b32_e32 v5, v1
+; GFX908-NEXT: v_mov_b32_e32 v4, v0
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: v_mov_b32_e32 v1, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB7_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_addk_i32 s8, 0x200
+; GFX8-NEXT: s_and_b32 s8, s8, -4
+; GFX8-NEXT: v_mov_b32_e32 v2, s8
+; GFX8-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_max_f16_e32 v3, v0, v0
+; GFX8-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX8-NEXT: v_and_b32_e32 v4, 0xffffff, v1
+; GFX8-NEXT: v_min_f16_sdwa v0, v0, v3 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX8-NEXT: v_mov_b32_e32 v5, v1
+; GFX8-NEXT: v_mov_b32_e32 v4, v0
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v1, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB7_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_addk_i32 s8, 0x200
+; GFX7-NEXT: s_and_b32 s8, s8, -4
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v0
+; GFX7-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffffff, v1
+; GFX7-NEXT: v_min_f32_e32 v0, v0, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX7-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX7-NEXT: v_mov_b32_e32 v5, v1
+; GFX7-NEXT: v_mov_b32_e32 v4, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v1, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB7_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_addk_i32 s8, 0x200
+; GFX6-NEXT: s_and_b32 s8, s8, -4
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v0
+; GFX6-NEXT: .LBB7_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffffff, v1
+; GFX6-NEXT: v_min_f32_e32 v0, v0, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX6-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX6-NEXT: v_mov_b32_e32 v5, v1
+; GFX6-NEXT: v_mov_b32_e32 v4, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v1, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB7_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(7) %ptr, i32 256
+ %unused = atomicrmw fmin ptr addrspace(7) %gep, half %val syncscope("agent") seq_cst
+ ret void
+}
+
+define half @buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset__waterfall(ptr addrspace(7) %ptr, half %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset__waterfall:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v4, 0x200, v4
+; GFX12-NEXT: s_mov_b32 s1, exec_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX12-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: buffer_load_b32 v6, v7, s[4:7], null offen
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB8_1
+; GFX12-NEXT: ; %bb.2:
+; GFX12-NEXT: s_mov_b32 exec_lo, s1
+; GFX12-NEXT: v_max_num_f16_e32 v8, v5, v5
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Loop Header: Depth=1
+; GFX12-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX12-NEXT: s_mov_b32 s2, exec_lo
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f16_e32 v4, v4, v4
+; GFX12-NEXT: v_min_num_f16_e32 v4, v4, v8
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_or_b32 v5, 0xffffff, v6, v4
+; GFX12-NEXT: v_mov_b32_e32 v4, v5
+; GFX12-NEXT: v_mov_b32_e32 v5, v6
+; GFX12-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX12-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB8_4
+; GFX12-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX12-NEXT: s_mov_b32 exec_lo, s2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX12-NEXT: v_mov_b32_e32 v6, v4
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB8_3
+; GFX12-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset__waterfall:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v4, 0x200, v4
+; GFX940-NEXT: v_and_b32_e32 v8, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v9, 0xffffff
+; GFX940-NEXT: s_mov_b64 s[2:3], exec
+; GFX940-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: buffer_load_dword v7, v8, s[4:7], 0 offen
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB8_1
+; GFX940-NEXT: ; %bb.2:
+; GFX940-NEXT: s_mov_b64 exec, s[2:3]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_max_f16_e32 v10, v5, v5
+; GFX940-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Loop Header: Depth=1
+; GFX940-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_max_f16_sdwa v4, v7, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX940-NEXT: v_min_f16_sdwa v4, v4, v10 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX940-NEXT: s_mov_b64 s[8:9], exec
+; GFX940-NEXT: v_and_or_b32 v6, v7, v9, v4
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[6:7]
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX940-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[4:7], 0 offen sc0
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB8_4
+; GFX940-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX940-NEXT: s_mov_b64 exec, s[8:9]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB8_3
+; GFX940-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset__waterfall:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v4, 0x200, v4
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX11-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: buffer_load_b32 v6, v7, s[4:7], 0 offen
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB8_1
+; GFX11-NEXT: ; %bb.2:
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: v_max_f16_e32 v8, v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Loop Header: Depth=1
+; GFX11-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f16_e32 v4, v4, v4
+; GFX11-NEXT: v_min_f16_e32 v4, v4, v8
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_or_b32 v5, 0xffffff, v6, v4
+; GFX11-NEXT: v_mov_b32_e32 v4, v5
+; GFX11-NEXT: v_mov_b32_e32 v5, v6
+; GFX11-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX11-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], 0 offen glc
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB8_4
+; GFX11-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX11-NEXT: v_mov_b32_e32 v6, v4
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB8_3
+; GFX11-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset__waterfall:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v4, 0x200, v4
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX10-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB8_1
+; GFX10-NEXT: ; %bb.2:
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: v_max_f16_e32 v8, v5, v5
+; GFX10-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Loop Header: Depth=1
+; GFX10-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_max_f16_sdwa v4, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_min_f16_e32 v4, v4, v8
+; GFX10-NEXT: v_and_b32_e32 v4, 0xffff, v4
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX10-NEXT: v_and_or_b32 v5, 0xffffff, v6, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v5
+; GFX10-NEXT: v_mov_b32_e32 v5, v6
+; GFX10-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX10-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB8_4
+; GFX10-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB8_3
+; GFX10-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset__waterfall:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v4, 0x200, v4
+; GFX90A-NEXT: v_and_b32_e32 v8, -4, v4
+; GFX90A-NEXT: v_mov_b32_e32 v9, 0xffffff
+; GFX90A-NEXT: s_mov_b64 s[6:7], exec
+; GFX90A-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_nop 0
+; GFX90A-NEXT: buffer_load_dword v7, v8, s[8:11], 0 offen
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB8_1
+; GFX90A-NEXT: ; %bb.2:
+; GFX90A-NEXT: s_mov_b64 exec, s[6:7]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_max_f16_e32 v10, v5, v5
+; GFX90A-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Loop Header: Depth=1
+; GFX90A-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_max_f16_sdwa v4, v7, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX90A-NEXT: v_min_f16_sdwa v4, v4, v10 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v9, v4
+; GFX90A-NEXT: s_mov_b64 s[12:13], exec
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[6:7] op_sel:[0,1]
+; GFX90A-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX90A-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[8:11], 0 offen glc
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB8_4
+; GFX90A-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX90A-NEXT: s_mov_b64 exec, s[12:13]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB8_3
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset__waterfall:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v4, 0x200, v4
+; GFX908-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX908-NEXT: v_mov_b32_e32 v8, 0xffffff
+; GFX908-NEXT: s_mov_b64 s[6:7], exec
+; GFX908-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_nop 0
+; GFX908-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_1
+; GFX908-NEXT: ; %bb.2:
+; GFX908-NEXT: s_mov_b64 exec, s[6:7]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_max_f16_e32 v9, v5, v5
+; GFX908-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Loop Header: Depth=1
+; GFX908-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_max_f16_sdwa v4, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX908-NEXT: v_min_f16_sdwa v4, v4, v9 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX908-NEXT: v_and_or_b32 v5, v6, v8, v4
+; GFX908-NEXT: v_mov_b32_e32 v4, v5
+; GFX908-NEXT: s_mov_b64 s[12:13], exec
+; GFX908-NEXT: v_mov_b32_e32 v5, v6
+; GFX908-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX908-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_4
+; GFX908-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX908-NEXT: s_mov_b64 exec, s[12:13]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB8_3
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset__waterfall:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x200, v4
+; GFX8-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX8-NEXT: s_mov_b64 s[6:7], exec
+; GFX8-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_nop 0
+; GFX8-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_1
+; GFX8-NEXT: ; %bb.2:
+; GFX8-NEXT: s_mov_b64 exec, s[6:7]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_max_f16_e32 v8, v5, v5
+; GFX8-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Loop Header: Depth=1
+; GFX8-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v4, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX8-NEXT: v_min_f16_sdwa v4, v4, v8 dst_sel:BYTE_3 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffffff, v6
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v4
+; GFX8-NEXT: v_mov_b32_e32 v4, v5
+; GFX8-NEXT: s_mov_b64 s[12:13], exec
+; GFX8-NEXT: v_mov_b32_e32 v5, v6
+; GFX8-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX8-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_4
+; GFX8-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX8-NEXT: s_mov_b64 exec, s[12:13]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB8_3
+; GFX8-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset__waterfall:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x200, v4
+; GFX7-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB8_1
+; GFX7-NEXT: ; %bb.2:
+; GFX7-NEXT: s_mov_b64 exec, s[6:7]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v5
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v8, v4
+; GFX7-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Loop Header: Depth=1
+; GFX7-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffffff, v6
+; GFX7-NEXT: s_mov_b64 s[12:13], exec
+; GFX7-NEXT: v_min_f32_e32 v4, v4, v8
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX7-NEXT: v_or_b32_e32 v5, v5, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: v_mov_b32_e32 v5, v6
+; GFX7-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX7-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB8_4
+; GFX7-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX7-NEXT: s_mov_b64 exec, s[12:13]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB8_3
+; GFX7-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset__waterfall:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x200, v4
+; GFX6-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB8_1
+; GFX6-NEXT: ; %bb.2:
+; GFX6-NEXT: s_mov_b64 exec, s[6:7]
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v5
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v8, v4
+; GFX6-NEXT: .LBB8_3: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Loop Header: Depth=1
+; GFX6-NEXT: ; Child Loop BB8_4 Depth 2
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffffff, v6
+; GFX6-NEXT: s_mov_b64 s[12:13], exec
+; GFX6-NEXT: v_min_f32_e32 v4, v4, v8
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX6-NEXT: v_or_b32_e32 v5, v5, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: v_mov_b32_e32 v5, v6
+; GFX6-NEXT: .LBB8_4: ; Parent Loop BB8_3 Depth=1
+; GFX6-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB8_4
+; GFX6-NEXT: ; %bb.5: ; in Loop: Header=BB8_3 Depth=1
+; GFX6-NEXT: s_mov_b64 exec, s[12:13]
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB8_3
+; GFX6-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr half, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmin ptr addrspace(7) %gep, half %val syncscope("agent") seq_cst
+ ret half %result
+}
+
+; --------------------------------------------------------------------
+; bfloat
+; --------------------------------------------------------------------
+
+define bfloat @buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset(ptr addrspace(7) inreg %ptr, bfloat %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_addk_co_i32 s4, 0x200
+; GFX12-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX12-NEXT: s_and_b32 s4, s4, -4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_mov_b32_e32 v4, s4
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: buffer_load_b32 v1, v4, s[0:3], null offen
+; GFX12-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX12-NEXT: v_min_num_f32_e32 v0, v0, v5
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v3, 0x400000, v0
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX12-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX12-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[2:3], v4, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v2
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB9_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_addk_i32 s4, 0x200
+; GFX940-NEXT: s_and_b32 s4, s4, -4
+; GFX940-NEXT: v_mov_b32_e32 v4, s4
+; GFX940-NEXT: buffer_load_dword v1, v4, s[0:3], 0 offen
+; GFX940-NEXT: s_mov_b32 s6, 0xffffff
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX940-NEXT: s_movk_i32 s7, 0x7fff
+; GFX940-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX940-NEXT: v_min_f32_e32 v0, v0, v5
+; GFX940-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v3, 0x400000, v0
+; GFX940-NEXT: v_add3_u32 v2, v2, v0, s7
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX940-NEXT: v_and_or_b32 v0, v1, s6, v0
+; GFX940-NEXT: v_mov_b64_e32 v[2:3], v[0:1]
+; GFX940-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB9_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_addk_i32 s4, 0x200
+; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX11-NEXT: s_and_b32 s4, s4, -4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v4, s4
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: buffer_load_b32 v1, v4, s[0:3], 0 offen
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT: v_min_f32_e32 v0, v0, v5
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v3, 0x400000, v0
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX11-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[2:3], v4, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v2
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB9_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_addk_i32 s8, 0x200
+; GFX10-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX10-NEXT: s_and_b32 s8, s8, -4
+; GFX10-NEXT: v_mov_b32_e32 v4, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX10-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX10-NEXT: v_min_f32_e32 v0, v0, v5
+; GFX10-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v3, 0x400000, v0
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX10-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX10-NEXT: v_mov_b32_e32 v3, v1
+; GFX10-NEXT: v_mov_b32_e32 v2, v0
+; GFX10-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v2
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB9_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: s_addk_i32 s8, 0x200
+; GFX90A-NEXT: s_and_b32 s8, s8, -4
+; GFX90A-NEXT: v_mov_b32_e32 v4, s8
+; GFX90A-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX90A-NEXT: s_mov_b32 s10, 0xffffff
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX90A-NEXT: s_movk_i32 s11, 0x7fff
+; GFX90A-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX90A-NEXT: v_min_f32_e32 v0, v0, v5
+; GFX90A-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v3, 0x400000, v0
+; GFX90A-NEXT: v_add3_u32 v2, v2, v0, s11
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX90A-NEXT: v_and_or_b32 v0, v1, s10, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB9_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: s_addk_i32 s8, 0x200
+; GFX908-NEXT: s_and_b32 s8, s8, -4
+; GFX908-NEXT: v_mov_b32_e32 v4, s8
+; GFX908-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX908-NEXT: s_mov_b32 s10, 0xffffff
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX908-NEXT: s_movk_i32 s11, 0x7fff
+; GFX908-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX908-NEXT: v_min_f32_e32 v0, v0, v5
+; GFX908-NEXT: v_bfe_u32 v2, v0, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v3, 0x400000, v0
+; GFX908-NEXT: v_add3_u32 v2, v2, v0, s11
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX908-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX908-NEXT: v_and_or_b32 v0, v1, s10, v0
+; GFX908-NEXT: v_mov_b32_e32 v3, v1
+; GFX908-NEXT: v_mov_b32_e32 v2, v0
+; GFX908-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: v_mov_b32_e32 v1, v2
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB9_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_addk_i32 s8, 0x200
+; GFX8-NEXT: s_and_b32 s8, s8, -4
+; GFX8-NEXT: v_mov_b32_e32 v4, s8
+; GFX8-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v0
+; GFX8-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: v_min_f32_e32 v0, v0, v5
+; GFX8-NEXT: v_bfe_u32 v3, v0, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v0
+; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
+; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v0, v3, v6, vcc
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: v_and_b32_e32 v2, 0xffffff, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX8-NEXT: v_or_b32_e32 v0, v2, v0
+; GFX8-NEXT: v_mov_b32_e32 v3, v1
+; GFX8-NEXT: v_mov_b32_e32 v2, v0
+; GFX8-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v1, v2
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB9_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_addk_i32 s8, 0x200
+; GFX7-NEXT: s_and_b32 s8, s8, -4
+; GFX7-NEXT: v_mov_b32_e32 v4, s8
+; GFX7-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v0
+; GFX7-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_min_f32_e32 v0, v0, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffffff, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX7-NEXT: v_or_b32_e32 v0, v2, v0
+; GFX7-NEXT: v_mov_b32_e32 v3, v1
+; GFX7-NEXT: v_mov_b32_e32 v2, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v1, v2
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB9_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_addk_i32 s8, 0x200
+; GFX6-NEXT: s_and_b32 s8, s8, -4
+; GFX6-NEXT: v_mov_b32_e32 v4, s8
+; GFX6-NEXT: buffer_load_dword v1, v4, s[4:7], 0 offen
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v0
+; GFX6-NEXT: .LBB9_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: v_min_f32_e32 v0, v0, v5
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffffff, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX6-NEXT: v_or_b32_e32 v0, v2, v0
+; GFX6-NEXT: v_mov_b32_e32 v3, v1
+; GFX6-NEXT: v_mov_b32_e32 v2, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[2:3], v4, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v1, v2
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB9_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmin ptr addrspace(7) %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+define void @buffer_fat_ptr_agent_atomic_fmin_noret_bf16__offset(ptr addrspace(7) inreg %ptr, bfloat %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: s_addk_co_i32 s4, 0x200
+; GFX12-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX12-NEXT: s_and_b32 s4, s4, -4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_mov_b32_e32 v2, s4
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: buffer_load_b32 v1, v2, s[0:3], null offen
+; GFX12-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX12-NEXT: v_min_num_f32_e32 v0, v0, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v4, v0, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX12-NEXT: v_add3_u32 v4, v4, v0, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX12-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v2, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v4
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB10_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: s_addk_i32 s4, 0x200
+; GFX940-NEXT: s_and_b32 s4, s4, -4
+; GFX940-NEXT: v_mov_b32_e32 v2, s4
+; GFX940-NEXT: buffer_load_dword v1, v2, s[0:3], 0 offen
+; GFX940-NEXT: s_mov_b32 s6, 0xffffff
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX940-NEXT: s_movk_i32 s7, 0x7fff
+; GFX940-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX940-NEXT: v_min_f32_e32 v0, v0, v3
+; GFX940-NEXT: v_bfe_u32 v4, v0, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX940-NEXT: v_add3_u32 v4, v4, v0, s7
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX940-NEXT: v_and_or_b32 v0, v1, s6, v0
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[0:1]
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB10_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: s_addk_i32 s4, 0x200
+; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX11-NEXT: s_and_b32 s4, s4, -4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_mov_b32_e32 v2, s4
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: buffer_load_b32 v1, v2, s[0:3], 0 offen
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT: v_min_f32_e32 v0, v0, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v4, v0, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT: v_add3_u32 v4, v4, v0, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX11-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v2, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v4
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB10_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_addk_i32 s8, 0x200
+; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX10-NEXT: s_and_b32 s8, s8, -4
+; GFX10-NEXT: v_mov_b32_e32 v2, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX10-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX10-NEXT: v_min_f32_e32 v0, v0, v3
+; GFX10-NEXT: v_bfe_u32 v4, v0, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX10-NEXT: v_add3_u32 v4, v4, v0, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX10-NEXT: v_and_or_b32 v0, 0xffffff, v1, v0
+; GFX10-NEXT: v_mov_b32_e32 v5, v1
+; GFX10-NEXT: v_mov_b32_e32 v4, v0
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v4
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB10_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: s_addk_i32 s8, 0x200
+; GFX90A-NEXT: s_and_b32 s8, s8, -4
+; GFX90A-NEXT: v_mov_b32_e32 v2, s8
+; GFX90A-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX90A-NEXT: s_mov_b32 s10, 0xffffff
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX90A-NEXT: s_movk_i32 s11, 0x7fff
+; GFX90A-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX90A-NEXT: v_min_f32_e32 v0, v0, v3
+; GFX90A-NEXT: v_bfe_u32 v4, v0, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX90A-NEXT: v_add3_u32 v4, v4, v0, s11
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX90A-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX90A-NEXT: v_and_or_b32 v0, v1, s10, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB10_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: s_addk_i32 s8, 0x200
+; GFX908-NEXT: s_and_b32 s8, s8, -4
+; GFX908-NEXT: v_mov_b32_e32 v2, s8
+; GFX908-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX908-NEXT: s_mov_b32 s10, 0xffffff
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX908-NEXT: s_movk_i32 s11, 0x7fff
+; GFX908-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX908-NEXT: v_min_f32_e32 v0, v0, v3
+; GFX908-NEXT: v_bfe_u32 v4, v0, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v5, 0x400000, v0
+; GFX908-NEXT: v_add3_u32 v4, v4, v0, s11
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX908-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX908-NEXT: v_and_or_b32 v0, v1, s10, v0
+; GFX908-NEXT: v_mov_b32_e32 v5, v1
+; GFX908-NEXT: v_mov_b32_e32 v4, v0
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: v_mov_b32_e32 v1, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB10_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_addk_i32 s8, 0x200
+; GFX8-NEXT: s_and_b32 s8, s8, -4
+; GFX8-NEXT: v_mov_b32_e32 v2, s8
+; GFX8-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v0
+; GFX8-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: v_min_f32_e32 v0, v0, v3
+; GFX8-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v0
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
+; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v0, v5, v6, vcc
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: v_and_b32_e32 v4, 0xffffff, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX8-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX8-NEXT: v_mov_b32_e32 v5, v1
+; GFX8-NEXT: v_mov_b32_e32 v4, v0
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v1, v4
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB10_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_addk_i32 s8, 0x200
+; GFX7-NEXT: s_and_b32 s8, s8, -4
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX7-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_min_f32_e32 v0, v0, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffffff, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX7-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX7-NEXT: v_mov_b32_e32 v5, v1
+; GFX7-NEXT: v_mov_b32_e32 v4, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: v_mov_b32_e32 v1, v4
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB10_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: s_addk_i32 s8, 0x200
+; GFX6-NEXT: s_and_b32 s8, s8, -4
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dword v1, v2, s[4:7], 0 offen
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX6-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: v_min_f32_e32 v0, v0, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffffff, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 24, v0
+; GFX6-NEXT: v_or_b32_e32 v0, v4, v0
+; GFX6-NEXT: v_mov_b32_e32 v5, v1
+; GFX6-NEXT: v_mov_b32_e32 v4, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v2, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: v_mov_b32_e32 v1, v4
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB10_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(7) %ptr, i32 256
+ %unused = atomicrmw fmin ptr addrspace(7) %gep, bfloat %val syncscope("agent") seq_cst
+ ret void
+}
+
+define bfloat @buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__waterfall(ptr addrspace(7) %ptr, bfloat %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__waterfall:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v4, 0x200, v4
+; GFX12-NEXT: s_mov_b32 s1, exec_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX12-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: buffer_load_b32 v6, v7, s[4:7], null offen
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB11_1
+; GFX12-NEXT: ; %bb.2:
+; GFX12-NEXT: s_mov_b32 exec_lo, s1
+; GFX12-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Loop Header: Depth=1
+; GFX12-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX12-NEXT: s_mov_b32 s2, exec_lo
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: v_min_num_f32_e32 v4, v4, v8
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v4
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX12-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e32 v4, v5, v9, vcc_lo
+; GFX12-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX12-NEXT: v_and_or_b32 v5, 0xffffff, v6, v4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_mov_b32_e32 v4, v5
+; GFX12-NEXT: v_mov_b32_e32 v5, v6
+; GFX12-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX12-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB11_4
+; GFX12-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX12-NEXT: s_mov_b32 exec_lo, s2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX12-NEXT: v_mov_b32_e32 v6, v4
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB11_3
+; GFX12-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__waterfall:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v4, 0x200, v4
+; GFX940-NEXT: v_and_b32_e32 v8, -4, v4
+; GFX940-NEXT: v_mov_b32_e32 v9, 0xffffff
+; GFX940-NEXT: s_mov_b64 s[2:3], exec
+; GFX940-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: buffer_load_dword v7, v8, s[4:7], 0 offen
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB11_1
+; GFX940-NEXT: ; %bb.2:
+; GFX940-NEXT: s_mov_b64 exec, s[2:3]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v10, 16, v5
+; GFX940-NEXT: s_movk_i32 s10, 0x7fff
+; GFX940-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Loop Header: Depth=1
+; GFX940-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, 24, v7
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX940-NEXT: v_min_f32_e32 v4, v4, v10
+; GFX940-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX940-NEXT: v_add3_u32 v5, v5, v4, s10
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_mov_b64 s[8:9], exec
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX940-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX940-NEXT: v_and_or_b32 v6, v7, v9, v4
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[6:7]
+; GFX940-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX940-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[4:7], 0 offen sc0
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB11_4
+; GFX940-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX940-NEXT: s_mov_b64 exec, s[8:9]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB11_3
+; GFX940-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__waterfall:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v4, 0x200, v4
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX11-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: buffer_load_b32 v6, v7, s[4:7], 0 offen
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB11_1
+; GFX11-NEXT: ; %bb.2:
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Loop Header: Depth=1
+; GFX11-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: v_min_f32_e32 v4, v4, v8
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v4
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e32 v4, v5, v9, vcc_lo
+; GFX11-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX11-NEXT: v_and_or_b32 v5, 0xffffff, v6, v4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_mov_b32_e32 v4, v5
+; GFX11-NEXT: v_mov_b32_e32 v5, v6
+; GFX11-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX11-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], 0 offen glc
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB11_4
+; GFX11-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX11-NEXT: v_mov_b32_e32 v6, v4
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB11_3
+; GFX11-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__waterfall:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v4, 0x200, v4
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX10-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB11_1
+; GFX10-NEXT: ; %bb.2:
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX10-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Loop Header: Depth=1
+; GFX10-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX10-NEXT: v_min_f32_e32 v4, v4, v8
+; GFX10-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v4
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX10-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v4, v5, v9, vcc_lo
+; GFX10-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX10-NEXT: v_and_or_b32 v5, 0xffffff, v6, v4
+; GFX10-NEXT: v_mov_b32_e32 v4, v5
+; GFX10-NEXT: v_mov_b32_e32 v5, v6
+; GFX10-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX10-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB11_4
+; GFX10-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB11_3
+; GFX10-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__waterfall:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v4, 0x200, v4
+; GFX90A-NEXT: v_and_b32_e32 v8, -4, v4
+; GFX90A-NEXT: v_mov_b32_e32 v9, 0xffffff
+; GFX90A-NEXT: s_mov_b64 s[6:7], exec
+; GFX90A-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_nop 0
+; GFX90A-NEXT: buffer_load_dword v7, v8, s[8:11], 0 offen
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_1
+; GFX90A-NEXT: ; %bb.2:
+; GFX90A-NEXT: s_mov_b64 exec, s[6:7]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v10, 16, v5
+; GFX90A-NEXT: s_movk_i32 s14, 0x7fff
+; GFX90A-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Loop Header: Depth=1
+; GFX90A-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, 24, v7
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX90A-NEXT: v_min_f32_e32 v4, v4, v10
+; GFX90A-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX90A-NEXT: v_add3_u32 v5, v5, v4, s14
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX90A-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX90A-NEXT: v_and_or_b32 v6, v7, v9, v4
+; GFX90A-NEXT: s_mov_b64 s[12:13], exec
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[6:7] op_sel:[0,1]
+; GFX90A-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX90A-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[8:11], 0 offen glc
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_4
+; GFX90A-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX90A-NEXT: s_mov_b64 exec, s[12:13]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB11_3
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__waterfall:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v4, 0x200, v4
+; GFX908-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX908-NEXT: v_mov_b32_e32 v8, 0xffffff
+; GFX908-NEXT: s_mov_b64 s[6:7], exec
+; GFX908-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_nop 0
+; GFX908-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_1
+; GFX908-NEXT: ; %bb.2:
+; GFX908-NEXT: s_mov_b64 exec, s[6:7]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v9, 16, v5
+; GFX908-NEXT: s_movk_i32 s14, 0x7fff
+; GFX908-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Loop Header: Depth=1
+; GFX908-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX908-NEXT: v_min_f32_e32 v4, v4, v9
+; GFX908-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX908-NEXT: v_add3_u32 v5, v5, v4, s14
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v4
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX908-NEXT: v_cndmask_b32_e32 v4, v5, v10, vcc
+; GFX908-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX908-NEXT: v_and_or_b32 v5, v6, v8, v4
+; GFX908-NEXT: v_mov_b32_e32 v4, v5
+; GFX908-NEXT: s_mov_b64 s[12:13], exec
+; GFX908-NEXT: v_mov_b32_e32 v5, v6
+; GFX908-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX908-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_4
+; GFX908-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX908-NEXT: s_mov_b64 exec, s[12:13]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB11_3
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__waterfall:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x200, v4
+; GFX8-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX8-NEXT: s_mov_b64 s[6:7], exec
+; GFX8-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_nop 0
+; GFX8-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_1
+; GFX8-NEXT: ; %bb.2:
+; GFX8-NEXT: s_mov_b64 exec, s[6:7]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX8-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Loop Header: Depth=1
+; GFX8-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX8-NEXT: v_min_f32_e32 v4, v4, v8
+; GFX8-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v4
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v4, v5, v9, vcc
+; GFX8-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffffff, v6
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v4
+; GFX8-NEXT: v_mov_b32_e32 v4, v5
+; GFX8-NEXT: s_mov_b64 s[12:13], exec
+; GFX8-NEXT: v_mov_b32_e32 v5, v6
+; GFX8-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX8-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_4
+; GFX8-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX8-NEXT: s_mov_b64 exec, s[12:13]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB11_3
+; GFX8-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__waterfall:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x200, v4
+; GFX7-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB11_1
+; GFX7-NEXT: ; %bb.2:
+; GFX7-NEXT: s_mov_b64 exec, s[6:7]
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v5
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v4
+; GFX7-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Loop Header: Depth=1
+; GFX7-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_min_f32_e32 v4, v4, v8
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffffff, v6
+; GFX7-NEXT: v_or_b32_e32 v5, v5, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, v5
+; GFX7-NEXT: s_mov_b64 s[12:13], exec
+; GFX7-NEXT: v_mov_b32_e32 v5, v6
+; GFX7-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX7-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB11_4
+; GFX7-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX7-NEXT: s_mov_b64 exec, s[12:13]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB11_3
+; GFX7-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__waterfall:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v4, vcc, 0x200, v4
+; GFX6-NEXT: v_and_b32_e32 v7, -4, v4
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: buffer_load_dword v6, v7, s[8:11], 0 offen
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB11_1
+; GFX6-NEXT: ; %bb.2:
+; GFX6-NEXT: s_mov_b64 exec, s[6:7]
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v5
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: v_and_b32_e32 v8, 0xffff0000, v4
+; GFX6-NEXT: .LBB11_3: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Loop Header: Depth=1
+; GFX6-NEXT: ; Child Loop BB11_4 Depth 2
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 24, v6
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_min_f32_e32 v4, v4, v8
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 24, v4
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffffff, v6
+; GFX6-NEXT: v_or_b32_e32 v5, v5, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, v5
+; GFX6-NEXT: s_mov_b64 s[12:13], exec
+; GFX6-NEXT: v_mov_b32_e32 v5, v6
+; GFX6-NEXT: .LBB11_4: ; Parent Loop BB11_3 Depth=1
+; GFX6-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB11_4
+; GFX6-NEXT: ; %bb.5: ; in Loop: Header=BB11_3 Depth=1
+; GFX6-NEXT: s_mov_b64 exec, s[12:13]
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB11_3
+; GFX6-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 24, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr bfloat, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmin ptr addrspace(7) %gep, bfloat %val syncscope("agent") seq_cst
+ ret bfloat %result
+}
+
+; --------------------------------------------------------------------
+; <2 x half>
+; --------------------------------------------------------------------
+
+define <2 x half> @buffer_fat_ptr_agent_atomic_fmin_ret_v2f16__offset(ptr addrspace(7) inreg %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s4
+; GFX12-NEXT: s_addk_co_i32 s4, 0x400
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_mov_b32_e32 v3, s4
+; GFX12-NEXT: v_pk_max_num_f16 v2, v1, v1
+; GFX12-NEXT: buffer_load_b32 v0, v0, s[0:3], null offen offset:1024
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v5, v0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_max_num_f16 v0, v5, v5
+; GFX12-NEXT: v_pk_min_num_f16 v4, v0, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v3, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB12_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, v0
+; GFX940-NEXT: v_mov_b32_e32 v0, s4
+; GFX940-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen offset:1024
+; GFX940-NEXT: s_add_i32 s6, s4, 0x400
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_pk_max_f16 v2, v1, v1
+; GFX940-NEXT: v_mov_b32_e32 v3, s6
+; GFX940-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v0
+; GFX940-NEXT: v_pk_max_f16 v0, v5, v5
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_pk_min_f16 v4, v0, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_mov_b64_e32 v[0:1], v[4:5]
+; GFX940-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB12_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s4
+; GFX11-NEXT: s_addk_i32 s4, 0x400
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_mov_b32_e32 v3, s4
+; GFX11-NEXT: v_pk_max_f16 v2, v1, v1
+; GFX11-NEXT: buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v5, v0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_max_f16 v0, v5, v5
+; GFX11-NEXT: v_pk_min_f16 v4, v0, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v3, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB12_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_mov_b32_e32 v0, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: v_mov_b32_e32 v3, s8
+; GFX10-NEXT: v_pk_max_f16 v2, v1, v1
+; GFX10-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v5, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_pk_max_f16 v0, v5, v5
+; GFX10-NEXT: v_pk_min_f16 v4, v0, v2
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB12_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: v_mov_b32_e32 v0, s8
+; GFX90A-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX90A-NEXT: s_add_i32 s10, s8, 0x400
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_pk_max_f16 v2, v1, v1
+; GFX90A-NEXT: v_mov_b32_e32 v3, s10
+; GFX90A-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v0
+; GFX90A-NEXT: v_pk_max_f16 v0, v5, v5
+; GFX90A-NEXT: v_pk_min_f16 v4, v0, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB12_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: v_mov_b32_e32 v0, s8
+; GFX908-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_add_i32 s10, s8, 0x400
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_pk_max_f16 v2, v1, v1
+; GFX908-NEXT: v_mov_b32_e32 v3, s10
+; GFX908-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v5, v0
+; GFX908-NEXT: v_pk_max_f16 v0, v5, v5
+; GFX908-NEXT: v_pk_min_f16 v4, v0, v2
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB12_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_mov_b32_e32 v0, s8
+; GFX8-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_add_i32 s10, s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_max_f16_sdwa v2, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v3, v1, v1
+; GFX8-NEXT: v_mov_b32_e32 v4, s10
+; GFX8-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_max_f16_sdwa v0, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v1, v6, v6
+; GFX8-NEXT: v_min_f16_sdwa v0, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v1, v1, v3
+; GFX8-NEXT: v_or_b32_e32 v5, v1, v0
+; GFX8-NEXT: v_mov_b32_e32 v0, v5
+; GFX8-NEXT: v_mov_b32_e32 v1, v6
+; GFX8-NEXT: buffer_atomic_cmpswap v[0:1], v4, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB12_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dword v3, v2, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v0
+; GFX7-NEXT: s_add_i32 s10, s8, 0x400
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, s10
+; GFX7-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v3
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v6
+; GFX7-NEXT: v_or_b32_e32 v6, v0, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v5
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v0
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v4, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB12_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dword v3, v2, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v0
+; GFX6-NEXT: s_add_i32 s10, s8, 0x400
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, s10
+; GFX6-NEXT: .LBB12_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v0
+; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v3
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v6
+; GFX6-NEXT: v_or_b32_e32 v6, v0, v1
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v5
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v0
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v4, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB12_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmin ptr addrspace(7) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+define void @buffer_fat_ptr_agent_atomic_fmin_noret_v2f16__offset(ptr addrspace(7) inreg %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_v2f16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v1, s4
+; GFX12-NEXT: s_addk_co_i32 s4, 0x400
+; GFX12-NEXT: v_pk_max_num_f16 v2, v0, v0
+; GFX12-NEXT: v_mov_b32_e32 v3, s4
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: buffer_load_b32 v1, v1, s[0:3], null offen offset:1024
+; GFX12-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v0, v1, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_min_num_f16 v0, v0, v2
+; GFX12-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v4
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB13_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_v2f16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, s4
+; GFX940-NEXT: buffer_load_dword v1, v1, s[0:3], 0 offen offset:1024
+; GFX940-NEXT: s_add_i32 s6, s4, 0x400
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_pk_max_f16 v2, v0, v0
+; GFX940-NEXT: v_mov_b32_e32 v3, s6
+; GFX940-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v0, v1, v1
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_pk_min_f16 v0, v0, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[0:1]
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB13_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_v2f16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v1, s4
+; GFX11-NEXT: s_addk_i32 s4, 0x400
+; GFX11-NEXT: v_pk_max_f16 v2, v0, v0
+; GFX11-NEXT: v_mov_b32_e32 v3, s4
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: buffer_load_b32 v1, v1, s[0:3], 0 offen offset:1024
+; GFX11-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v0, v1, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_min_f16 v0, v0, v2
+; GFX11-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v4
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB13_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_v2f16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: v_pk_max_f16 v2, v0, v0
+; GFX10-NEXT: v_mov_b32_e32 v3, s8
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v0, v1, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_pk_min_f16 v0, v0, v2
+; GFX10-NEXT: v_mov_b32_e32 v5, v1
+; GFX10-NEXT: v_mov_b32_e32 v4, v0
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v4
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB13_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_v2f16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, s8
+; GFX90A-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX90A-NEXT: s_add_i32 s10, s8, 0x400
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_pk_max_f16 v2, v0, v0
+; GFX90A-NEXT: v_mov_b32_e32 v3, s10
+; GFX90A-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v0, v1, v1
+; GFX90A-NEXT: v_pk_min_f16 v0, v0, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB13_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_v2f16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, s8
+; GFX908-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_add_i32 s10, s8, 0x400
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_pk_max_f16 v2, v0, v0
+; GFX908-NEXT: v_mov_b32_e32 v3, s10
+; GFX908-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v0, v1, v1
+; GFX908-NEXT: v_pk_min_f16 v0, v0, v2
+; GFX908-NEXT: v_mov_b32_e32 v5, v1
+; GFX908-NEXT: v_mov_b32_e32 v4, v0
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v3, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v1
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: v_mov_b32_e32 v1, v4
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB13_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_v2f16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, s8
+; GFX8-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_add_i32 s10, s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_max_f16_sdwa v2, v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v3, v0, v0
+; GFX8-NEXT: v_mov_b32_e32 v4, s10
+; GFX8-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v0, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v1, v1
+; GFX8-NEXT: v_min_f16_sdwa v0, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v3
+; GFX8-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX8-NEXT: v_mov_b32_e32 v6, v1
+; GFX8-NEXT: v_mov_b32_e32 v5, v0
+; GFX8-NEXT: buffer_atomic_cmpswap v[5:6], v4, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v1
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: v_mov_b32_e32 v1, v5
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB13_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_v2f16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dword v2, v2, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v0
+; GFX7-NEXT: s_add_i32 s10, s8, 0x400
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v2
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v1
+; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v5
+; GFX7-NEXT: v_mov_b32_e32 v2, s10
+; GFX7-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_min_f32_e32 v5, v5, v0
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v1
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_or_b32_e32 v5, v3, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX7-NEXT: v_or_b32_e32 v4, v6, v3
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v2, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB13_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_v2f16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dword v2, v2, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v0
+; GFX6-NEXT: s_add_i32 s10, s8, 0x400
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v2
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v1
+; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v5
+; GFX6-NEXT: v_mov_b32_e32 v2, s10
+; GFX6-NEXT: .LBB13_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v4
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v3
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_min_f32_e32 v5, v5, v0
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v1
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_or_b32_e32 v5, v3, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v7
+; GFX6-NEXT: v_or_b32_e32 v4, v6, v3
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v2, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v6
+; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v6
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB13_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(7) %ptr, i32 256
+ %unused = atomicrmw fmin ptr addrspace(7) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x half> @buffer_fat_ptr_agent_atomic_fmin_ret_v2f16__offset__waterfall(ptr addrspace(7) %ptr, <2 x half> %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2f16__offset__waterfall:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX12-NEXT: s_mov_b32 s1, exec_lo
+; GFX12-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: buffer_load_b32 v6, v4, s[4:7], null offen offset:1024
+; GFX12-NEXT: ; implicit-def: $vgpr4
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB14_1
+; GFX12-NEXT: ; %bb.2:
+; GFX12-NEXT: s_mov_b32 exec_lo, s1
+; GFX12-NEXT: v_pk_max_num_f16 v8, v5, v5
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Loop Header: Depth=1
+; GFX12-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_pk_max_num_f16 v4, v6, v6
+; GFX12-NEXT: s_mov_b32 s2, exec_lo
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_pk_min_num_f16 v5, v4, v8
+; GFX12-NEXT: v_mov_b32_e32 v4, v5
+; GFX12-NEXT: v_mov_b32_e32 v5, v6
+; GFX12-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX12-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB14_4
+; GFX12-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX12-NEXT: s_mov_b32 exec_lo, s2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX12-NEXT: v_mov_b32_e32 v6, v4
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB14_3
+; GFX12-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2f16__offset__waterfall:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v8, 0x400, v4
+; GFX940-NEXT: s_mov_b64 s[2:3], exec
+; GFX940-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: buffer_load_dword v7, v4, s[4:7], 0 offen offset:1024
+; GFX940-NEXT: ; implicit-def: $vgpr4
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB14_1
+; GFX940-NEXT: ; %bb.2:
+; GFX940-NEXT: s_mov_b64 exec, s[2:3]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_pk_max_f16 v9, v5, v5
+; GFX940-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Loop Header: Depth=1
+; GFX940-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_pk_max_f16 v4, v7, v7
+; GFX940-NEXT: s_mov_b64 s[8:9], exec
+; GFX940-NEXT: v_pk_min_f16 v6, v4, v9
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[6:7]
+; GFX940-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX940-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[4:7], 0 offen sc0
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB14_4
+; GFX940-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX940-NEXT: s_mov_b64 exec, s[8:9]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB14_3
+; GFX940-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2f16__offset__waterfall:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: buffer_load_b32 v6, v4, s[4:7], 0 offen offset:1024
+; GFX11-NEXT: ; implicit-def: $vgpr4
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB14_1
+; GFX11-NEXT: ; %bb.2:
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: v_pk_max_f16 v8, v5, v5
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Loop Header: Depth=1
+; GFX11-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_pk_max_f16 v4, v6, v6
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_pk_min_f16 v5, v4, v8
+; GFX11-NEXT: v_mov_b32_e32 v4, v5
+; GFX11-NEXT: v_mov_b32_e32 v5, v6
+; GFX11-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX11-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], 0 offen glc
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB14_4
+; GFX11-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX11-NEXT: v_mov_b32_e32 v6, v4
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB14_3
+; GFX11-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2f16__offset__waterfall:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX10-NEXT: ; implicit-def: $vgpr4
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB14_1
+; GFX10-NEXT: ; %bb.2:
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: v_pk_max_f16 v8, v5, v5
+; GFX10-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Loop Header: Depth=1
+; GFX10-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_pk_max_f16 v4, v6, v6
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_pk_min_f16 v5, v4, v8
+; GFX10-NEXT: v_mov_b32_e32 v4, v5
+; GFX10-NEXT: v_mov_b32_e32 v5, v6
+; GFX10-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX10-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB14_4
+; GFX10-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB14_3
+; GFX10-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2f16__offset__waterfall:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v8, 0x400, v4
+; GFX90A-NEXT: s_mov_b64 s[6:7], exec
+; GFX90A-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_nop 0
+; GFX90A-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen offset:1024
+; GFX90A-NEXT: ; implicit-def: $vgpr4
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_1
+; GFX90A-NEXT: ; %bb.2:
+; GFX90A-NEXT: s_mov_b64 exec, s[6:7]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_pk_max_f16 v9, v5, v5
+; GFX90A-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Loop Header: Depth=1
+; GFX90A-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_max_f16 v4, v7, v7
+; GFX90A-NEXT: v_pk_min_f16 v6, v4, v9
+; GFX90A-NEXT: s_mov_b64 s[12:13], exec
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[6:7] op_sel:[0,1]
+; GFX90A-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX90A-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[8:11], 0 offen glc
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_4
+; GFX90A-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX90A-NEXT: s_mov_b64 exec, s[12:13]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB14_3
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2f16__offset__waterfall:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v7, 0x400, v4
+; GFX908-NEXT: s_mov_b64 s[6:7], exec
+; GFX908-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_nop 0
+; GFX908-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX908-NEXT: ; implicit-def: $vgpr4
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_1
+; GFX908-NEXT: ; %bb.2:
+; GFX908-NEXT: s_mov_b64 exec, s[6:7]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_pk_max_f16 v8, v5, v5
+; GFX908-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Loop Header: Depth=1
+; GFX908-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_pk_max_f16 v4, v6, v6
+; GFX908-NEXT: v_pk_min_f16 v5, v4, v8
+; GFX908-NEXT: v_mov_b32_e32 v4, v5
+; GFX908-NEXT: s_mov_b64 s[12:13], exec
+; GFX908-NEXT: v_mov_b32_e32 v5, v6
+; GFX908-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX908-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_4
+; GFX908-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX908-NEXT: s_mov_b64 exec, s[12:13]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB14_3
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2f16__offset__waterfall:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x400, v4
+; GFX8-NEXT: s_mov_b64 s[6:7], exec
+; GFX8-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_nop 0
+; GFX8-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX8-NEXT: ; implicit-def: $vgpr4
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_1
+; GFX8-NEXT: ; %bb.2:
+; GFX8-NEXT: s_mov_b64 exec, s[6:7]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_max_f16_sdwa v8, v5, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v9, v5, v5
+; GFX8-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Loop Header: Depth=1
+; GFX8-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_max_f16_sdwa v4, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX8-NEXT: v_max_f16_e32 v5, v6, v6
+; GFX8-NEXT: v_min_f16_sdwa v4, v4, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX8-NEXT: v_min_f16_e32 v5, v5, v9
+; GFX8-NEXT: v_or_b32_e32 v5, v5, v4
+; GFX8-NEXT: v_mov_b32_e32 v4, v5
+; GFX8-NEXT: s_mov_b64 s[12:13], exec
+; GFX8-NEXT: v_mov_b32_e32 v5, v6
+; GFX8-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX8-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_4
+; GFX8-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX8-NEXT: s_mov_b64 exec, s[12:13]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB14_3
+; GFX8-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2f16__offset__waterfall:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v9, vcc, 0x400, v4
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen offset:1024
+; GFX7-NEXT: ; implicit-def: $vgpr4
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_1
+; GFX7-NEXT: ; %bb.2:
+; GFX7-NEXT: s_mov_b64 exec, s[6:7]
+; GFX7-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v5
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v10, v6
+; GFX7-NEXT: v_cvt_f32_f16_e32 v11, v8
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Loop Header: Depth=1
+; GFX7-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX7-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT: s_mov_b64 s[12:13], exec
+; GFX7-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX7-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v10
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v11
+; GFX7-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX7-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX7-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX7-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX7-NEXT: v_mov_b32_e32 v8, v6
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX7-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_atomic_cmpswap v[7:8], v9, s[8:11], 0 offen glc
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_4
+; GFX7-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX7-NEXT: s_mov_b64 exec, s[12:13]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX7-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB14_3
+; GFX7-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v0, v4
+; GFX7-NEXT: v_mov_b32_e32 v1, v5
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2f16__offset__waterfall:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v9, vcc, 0x400, v4
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: .LBB14_1: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen offset:1024
+; GFX6-NEXT: ; implicit-def: $vgpr4
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB14_1
+; GFX6-NEXT: ; %bb.2:
+; GFX6-NEXT: s_mov_b64 exec, s[6:7]
+; GFX6-NEXT: v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v5
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cvt_f32_f16_e32 v10, v6
+; GFX6-NEXT: v_cvt_f32_f16_e32 v11, v8
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: .LBB14_3: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Loop Header: Depth=1
+; GFX6-NEXT: ; Child Loop BB14_4 Depth 2
+; GFX6-NEXT: v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT: s_mov_b64 s[12:13], exec
+; GFX6-NEXT: v_cvt_f32_f16_e32 v6, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_cvt_f32_f16_e32 v7, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v10
+; GFX6-NEXT: v_min_f32_e32 v7, v7, v11
+; GFX6-NEXT: v_cvt_f16_f32_e32 v8, v6
+; GFX6-NEXT: v_cvt_f16_f32_e32 v7, v7
+; GFX6-NEXT: v_or_b32_e32 v6, v4, v5
+; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v8
+; GFX6-NEXT: v_or_b32_e32 v5, v7, v4
+; GFX6-NEXT: v_mov_b32_e32 v8, v6
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: .LBB14_4: ; Parent Loop BB14_3 Depth=1
+; GFX6-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: buffer_atomic_cmpswap v[7:8], v9, s[8:11], 0 offen glc
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB14_4
+; GFX6-NEXT: ; %bb.5: ; in Loop: Header=BB14_3 Depth=1
+; GFX6-NEXT: s_mov_b64 exec, s[12:13]
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v7
+; GFX6-NEXT: v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v7, v6
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB14_3
+; GFX6-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v0, v4
+; GFX6-NEXT: v_mov_b32_e32 v1, v5
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x half>, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmin ptr addrspace(7) %gep, <2 x half> %val syncscope("agent") seq_cst
+ ret <2 x half> %result
+}
+
+; --------------------------------------------------------------------
+; <2 x bfloat>
+; --------------------------------------------------------------------
+
+define <2 x bfloat> @buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset(ptr addrspace(7) inreg %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s4
+; GFX12-NEXT: s_addk_co_i32 s4, 0x400
+; GFX12-NEXT: s_mov_b32 s5, 0
+; GFX12-NEXT: v_mov_b32_e32 v4, s4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX12-NEXT: buffer_load_b32 v0, v0, s[0:3], null offen offset:1024
+; GFX12-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX12-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v6, v0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX12-NEXT: v_min_num_f32_e32 v1, v1, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v7, v1, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX12-NEXT: v_add3_u32 v7, v7, v1, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_cndmask_b32 v1, v7, v9 :: v_dual_lshlrev_b32 v0, 16, v6
+; GFX12-NEXT: v_min_num_f32_e32 v0, v0, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX12-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX12-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_cndmask_b32_e64 v0, v5, v8, s4
+; GFX12-NEXT: v_perm_b32 v5, v1, v0, 0x7060302
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v0, v5 :: v_dual_mov_b32 v1, v6
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v4, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX12-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5
+; GFX12-NEXT: s_cbranch_execnz .LBB15_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, v0
+; GFX940-NEXT: v_mov_b32_e32 v0, s4
+; GFX940-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen offset:1024
+; GFX940-NEXT: s_addk_i32 s4, 0x400
+; GFX940-NEXT: s_mov_b64 s[6:7], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX940-NEXT: s_movk_i32 s8, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX940-NEXT: s_mov_b32 s9, 0x7060302
+; GFX940-NEXT: v_mov_b32_e32 v4, s4
+; GFX940-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v7, v0
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 16, v7
+; GFX940-NEXT: v_and_b32_e32 v1, 0xffff0000, v7
+; GFX940-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX940-NEXT: v_min_f32_e32 v1, v1, v3
+; GFX940-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v1, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v0
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX940-NEXT: v_add3_u32 v5, v5, v0, s8
+; GFX940-NEXT: v_add3_u32 v8, v8, v1, s8
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX940-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_cndmask_b32_e32 v1, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v0, v5, v6, s[4:5]
+; GFX940-NEXT: v_perm_b32 v6, v1, v0, s9
+; GFX940-NEXT: v_mov_b64_e32 v[0:1], v[6:7]
+; GFX940-NEXT: buffer_atomic_cmpswap v[0:1], v4, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v7
+; GFX940-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX940-NEXT: s_cbranch_execnz .LBB15_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s4
+; GFX11-NEXT: s_addk_i32 s4, 0x400
+; GFX11-NEXT: s_mov_b32 s5, 0
+; GFX11-NEXT: v_mov_b32_e32 v4, s4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX11-NEXT: buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
+; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v6, v0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX11-NEXT: v_min_f32_e32 v1, v1, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v7, v1, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT: v_add3_u32 v7, v7, v1, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_cndmask_b32 v1, v7, v9 :: v_dual_lshlrev_b32 v0, 16, v6
+; GFX11-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX11-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX11-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_cndmask_b32_e64 v0, v5, v8, s4
+; GFX11-NEXT: v_perm_b32 v5, v1, v0, 0x7060302
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v0, v5 :: v_dual_mov_b32 v1, v6
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v4, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX11-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5
+; GFX11-NEXT: s_cbranch_execnz .LBB15_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_mov_b32_e32 v0, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: s_mov_b32 s9, 0
+; GFX10-NEXT: v_mov_b32_e32 v4, s8
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX10-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX10-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v6, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX10-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX10-NEXT: v_min_f32_e32 v1, v1, v3
+; GFX10-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v1, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX10-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v1, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s8, v0, v0
+; GFX10-NEXT: v_cndmask_b32_e32 v1, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v0, v5, v8, s8
+; GFX10-NEXT: v_perm_b32 v5, v1, v0, 0x7060302
+; GFX10-NEXT: v_mov_b32_e32 v0, v5
+; GFX10-NEXT: v_mov_b32_e32 v1, v6
+; GFX10-NEXT: buffer_atomic_cmpswap v[0:1], v4, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v6
+; GFX10-NEXT: s_or_b32 s9, vcc_lo, s9
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s9
+; GFX10-NEXT: s_cbranch_execnz .LBB15_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s9
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: v_mov_b32_e32 v0, s8
+; GFX90A-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX90A-NEXT: s_addk_i32 s8, 0x400
+; GFX90A-NEXT: s_mov_b64 s[10:11], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX90A-NEXT: s_movk_i32 s12, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX90A-NEXT: s_mov_b32 s13, 0x7060302
+; GFX90A-NEXT: v_mov_b32_e32 v4, s8
+; GFX90A-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v7, v0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v7
+; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v7
+; GFX90A-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX90A-NEXT: v_min_f32_e32 v1, v1, v3
+; GFX90A-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v1, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v0
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX90A-NEXT: v_add3_u32 v5, v5, v0, s12
+; GFX90A-NEXT: v_add3_u32 v8, v8, v1, s12
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[8:9], v0, v0
+; GFX90A-NEXT: v_cndmask_b32_e64 v0, v5, v6, s[8:9]
+; GFX90A-NEXT: v_cndmask_b32_e32 v1, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v1, v0, s13
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[6:7] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[0:1], v4, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v7
+; GFX90A-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX90A-NEXT: s_cbranch_execnz .LBB15_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: v_mov_b32_e32 v0, s8
+; GFX908-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_addk_i32 s8, 0x400
+; GFX908-NEXT: s_mov_b64 s[10:11], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX908-NEXT: s_movk_i32 s12, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX908-NEXT: s_mov_b32 s13, 0x7060302
+; GFX908-NEXT: v_mov_b32_e32 v4, s8
+; GFX908-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v6, v0
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX908-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX908-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX908-NEXT: v_min_f32_e32 v1, v1, v3
+; GFX908-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v1, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX908-NEXT: v_add3_u32 v5, v5, v0, s12
+; GFX908-NEXT: v_add3_u32 v8, v8, v1, s12
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX908-NEXT: v_cmp_u_f32_e64 s[8:9], v0, v0
+; GFX908-NEXT: v_cndmask_b32_e64 v0, v5, v7, s[8:9]
+; GFX908-NEXT: v_cndmask_b32_e32 v1, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v5, v1, v0, s13
+; GFX908-NEXT: v_mov_b32_e32 v0, v5
+; GFX908-NEXT: v_mov_b32_e32 v1, v6
+; GFX908-NEXT: buffer_atomic_cmpswap v[0:1], v4, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX908-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX908-NEXT: s_cbranch_execnz .LBB15_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_mov_b32_e32 v0, s8
+; GFX8-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_addk_i32 s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[10:11], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX8-NEXT: v_mov_b32_e32 v4, s8
+; GFX8-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v6, v0
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v6
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v6
+; GFX8-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX8-NEXT: v_min_f32_e32 v1, v1, v3
+; GFX8-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v1, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v0
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[8:9], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v1, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v5, v7, s[8:9]
+; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX8-NEXT: v_alignbit_b32 v5, v1, v0, 16
+; GFX8-NEXT: v_mov_b32_e32 v0, v5
+; GFX8-NEXT: v_mov_b32_e32 v1, v6
+; GFX8-NEXT: buffer_atomic_cmpswap v[0:1], v4, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v6
+; GFX8-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX8-NEXT: s_cbranch_execnz .LBB15_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dword v4, v2, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: s_addk_i32 s8, 0x400
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: s_mov_b64 s[10:11], 0
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v4
+; GFX7-NEXT: v_mov_b32_e32 v4, s8
+; GFX7-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v3
+; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v5
+; GFX7-NEXT: v_alignbit_b32 v0, v0, v6, 16
+; GFX7-NEXT: v_mov_b32_e32 v6, v1
+; GFX7-NEXT: v_mov_b32_e32 v5, v0
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v4, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v5
+; GFX7-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX7-NEXT: s_cbranch_execnz .LBB15_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dword v4, v2, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: s_addk_i32 s8, 0x400
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: s_mov_b64 s[10:11], 0
+; GFX6-NEXT: v_and_b32_e32 v2, 0xffff0000, v1
+; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v4
+; GFX6-NEXT: v_mov_b32_e32 v4, s8
+; GFX6-NEXT: .LBB15_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v0
+; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v1
+; GFX6-NEXT: v_min_f32_e32 v5, v5, v2
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v3
+; GFX6-NEXT: v_alignbit_b32 v1, v1, v0, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v0, 16, v5
+; GFX6-NEXT: v_alignbit_b32 v0, v0, v6, 16
+; GFX6-NEXT: v_mov_b32_e32 v6, v1
+; GFX6-NEXT: v_mov_b32_e32 v5, v0
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v4, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v1
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v5
+; GFX6-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX6-NEXT: v_lshlrev_b32_e32 v0, 16, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX6-NEXT: s_cbranch_execnz .LBB15_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmin ptr addrspace(7) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+define void @buffer_fat_ptr_agent_atomic_fmin_noret_v2bf16__offset(ptr addrspace(7) inreg %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_v2bf16__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v1, s4 :: v_dual_lshlrev_b32 v2, 16, v0
+; GFX12-NEXT: s_addk_co_i32 s4, 0x400
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_and_b32 v3, 0xffff0000, v0
+; GFX12-NEXT: buffer_load_b32 v1, v1, s[0:3], null offen offset:1024
+; GFX12-NEXT: s_mov_b32 s5, 0
+; GFX12-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX12-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_min_num_f32 v5, v5, v3 :: v_dual_min_num_f32 v0, v0, v2
+; GFX12-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX12-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX12-NEXT: v_add3_u32 v6, v6, v0, 0x7fff
+; GFX12-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX12-NEXT: v_cndmask_b32_e64 v0, v6, v8, s4
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v0, v5, v0, 0x7060302
+; GFX12-NEXT: v_dual_mov_b32 v6, v1 :: v_dual_mov_b32 v5, v0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[5:6], v4, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v1
+; GFX12-NEXT: v_mov_b32_e32 v1, v5
+; GFX12-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5
+; GFX12-NEXT: s_cbranch_execnz .LBB16_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_v2bf16__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, s4
+; GFX940-NEXT: buffer_load_dword v1, v1, s[0:3], 0 offen offset:1024
+; GFX940-NEXT: s_addk_i32 s4, 0x400
+; GFX940-NEXT: s_mov_b64 s[6:7], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v2, 16, v0
+; GFX940-NEXT: s_movk_i32 s8, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX940-NEXT: s_mov_b32 s9, 0x7060302
+; GFX940-NEXT: v_mov_b32_e32 v4, s4
+; GFX940-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX940-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v3
+; GFX940-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX940-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX940-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX940-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX940-NEXT: v_add3_u32 v6, v6, v0, s8
+; GFX940-NEXT: v_add3_u32 v8, v8, v5, s8
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: v_cmp_u_f32_e64 s[4:5], v0, v0
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX940-NEXT: v_cndmask_b32_e64 v0, v6, v7, s[4:5]
+; GFX940-NEXT: v_perm_b32 v0, v5, v0, s9
+; GFX940-NEXT: v_mov_b64_e32 v[6:7], v[0:1]
+; GFX940-NEXT: buffer_atomic_cmpswap v[6:7], v4, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v6, v1
+; GFX940-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX940-NEXT: v_mov_b32_e32 v1, v6
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX940-NEXT: s_cbranch_execnz .LBB16_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_v2bf16__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v1, s4 :: v_dual_lshlrev_b32 v2, 16, v0
+; GFX11-NEXT: s_addk_i32 s4, 0x400
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_and_b32 v3, 0xffff0000, v0
+; GFX11-NEXT: buffer_load_b32 v1, v1, s[0:3], 0 offen offset:1024
+; GFX11-NEXT: s_mov_b32 s5, 0
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_min_f32 v5, v5, v3 :: v_dual_min_f32 v0, v0, v2
+; GFX11-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX11-NEXT: v_add3_u32 v6, v6, v0, 0x7fff
+; GFX11-NEXT: v_cmp_u_f32_e64 s4, v0, v0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX11-NEXT: v_cndmask_b32_e64 v0, v6, v8, s4
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v0, v5, v0, 0x7060302
+; GFX11-NEXT: v_dual_mov_b32 v6, v1 :: v_dual_mov_b32 v5, v0
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[5:6], v4, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v1
+; GFX11-NEXT: v_mov_b32_e32 v1, v5
+; GFX11-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5
+; GFX11-NEXT: s_cbranch_execnz .LBB16_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_v2bf16__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v0
+; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX10-NEXT: v_mov_b32_e32 v4, s8
+; GFX10-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: s_mov_b32 s9, 0
+; GFX10-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v3
+; GFX10-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_add3_u32 v6, v6, v0, 0x7fff
+; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff
+; GFX10-NEXT: v_cmp_u_f32_e64 s8, v0, v0
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo
+; GFX10-NEXT: v_cndmask_b32_e64 v0, v6, v8, s8
+; GFX10-NEXT: v_perm_b32 v0, v5, v0, 0x7060302
+; GFX10-NEXT: v_mov_b32_e32 v6, v1
+; GFX10-NEXT: v_mov_b32_e32 v5, v0
+; GFX10-NEXT: buffer_atomic_cmpswap v[5:6], v4, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v5, v1
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: s_or_b32 s9, vcc_lo, s9
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s9
+; GFX10-NEXT: s_cbranch_execnz .LBB16_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s9
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_v2bf16__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, s8
+; GFX90A-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX90A-NEXT: s_addk_i32 s8, 0x400
+; GFX90A-NEXT: s_mov_b64 s[10:11], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v0
+; GFX90A-NEXT: s_movk_i32 s12, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX90A-NEXT: s_mov_b32 s13, 0x7060302
+; GFX90A-NEXT: v_mov_b32_e32 v4, s8
+; GFX90A-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX90A-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v3
+; GFX90A-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX90A-NEXT: v_add3_u32 v6, v6, v0, s12
+; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s12
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cmp_u_f32_e64 s[8:9], v0, v0
+; GFX90A-NEXT: v_cndmask_b32_e64 v0, v6, v7, s[8:9]
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX90A-NEXT: v_perm_b32 v0, v5, v0, s13
+; GFX90A-NEXT: v_pk_mov_b32 v[6:7], v[0:1], v[0:1] op_sel:[0,1]
+; GFX90A-NEXT: buffer_atomic_cmpswap v[6:7], v4, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v6, v1
+; GFX90A-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v6
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX90A-NEXT: s_cbranch_execnz .LBB16_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_v2bf16__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, s8
+; GFX908-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_addk_i32 s8, 0x400
+; GFX908-NEXT: s_mov_b64 s[10:11], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v0
+; GFX908-NEXT: s_movk_i32 s12, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX908-NEXT: s_mov_b32 s13, 0x7060302
+; GFX908-NEXT: v_mov_b32_e32 v4, s8
+; GFX908-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX908-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v3
+; GFX908-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX908-NEXT: v_add3_u32 v6, v6, v0, s12
+; GFX908-NEXT: v_add3_u32 v8, v8, v5, s12
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cmp_u_f32_e64 s[8:9], v0, v0
+; GFX908-NEXT: v_cndmask_b32_e64 v0, v6, v7, s[8:9]
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX908-NEXT: v_perm_b32 v0, v5, v0, s13
+; GFX908-NEXT: v_mov_b32_e32 v6, v1
+; GFX908-NEXT: v_mov_b32_e32 v5, v0
+; GFX908-NEXT: buffer_atomic_cmpswap v[5:6], v4, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v5, v1
+; GFX908-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX908-NEXT: s_cbranch_execnz .LBB16_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_v2bf16__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, s8
+; GFX8-NEXT: buffer_load_dword v1, v1, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_addk_i32 s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[10:11], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v0
+; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
+; GFX8-NEXT: v_mov_b32_e32 v4, s8
+; GFX8-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v1
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
+; GFX8-NEXT: v_min_f32_e32 v0, v0, v2
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v3
+; GFX8-NEXT: v_bfe_u32 v6, v0, 16, 1
+; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v0
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5
+; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6
+; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8
+; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v0
+; GFX8-NEXT: v_cmp_u_f32_e64 s[8:9], v0, v0
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc
+; GFX8-NEXT: v_cndmask_b32_e64 v0, v6, v7, s[8:9]
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v0, v5, v0, 16
+; GFX8-NEXT: v_mov_b32_e32 v6, v1
+; GFX8-NEXT: v_mov_b32_e32 v5, v0
+; GFX8-NEXT: buffer_atomic_cmpswap v[5:6], v4, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v5, v1
+; GFX8-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX8-NEXT: v_mov_b32_e32 v1, v5
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX8-NEXT: s_cbranch_execnz .LBB16_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_v2bf16__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: buffer_load_dword v2, v2, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: s_addk_i32 s8, 0x400
+; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v0
+; GFX7-NEXT: s_mov_b64 s[10:11], 0
+; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v2
+; GFX7-NEXT: v_mov_b32_e32 v2, s8
+; GFX7-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_min_f32_e32 v5, v5, v0
+; GFX7-NEXT: v_min_f32_e32 v6, v6, v1
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v3, 16
+; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: v_alignbit_b32 v3, v3, v6, 16
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: v_mov_b32_e32 v5, v3
+; GFX7-NEXT: buffer_atomic_cmpswap v[5:6], v2, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX7-NEXT: s_cbranch_execnz .LBB16_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmin_noret_v2bf16__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: buffer_load_dword v2, v2, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: s_addk_i32 s8, 0x400
+; GFX6-NEXT: v_mul_f32_e32 v1, 1.0, v1
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v0
+; GFX6-NEXT: s_mov_b64 s[10:11], 0
+; GFX6-NEXT: v_and_b32_e32 v0, 0xffff0000, v1
+; GFX6-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v2
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v2
+; GFX6-NEXT: v_mov_b32_e32 v2, s8
+; GFX6-NEXT: .LBB16_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_mul_f32_e32 v4, 1.0, v4
+; GFX6-NEXT: v_mul_f32_e32 v3, 1.0, v3
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v5, 0xffff0000, v4
+; GFX6-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_min_f32_e32 v5, v5, v0
+; GFX6-NEXT: v_min_f32_e32 v6, v6, v1
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v3, 16
+; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: v_alignbit_b32 v3, v3, v6, 16
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: v_mov_b32_e32 v5, v3
+; GFX6-NEXT: buffer_atomic_cmpswap v[5:6], v2, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v5, v4
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX6-NEXT: s_or_b64 s[10:11], vcc, s[10:11]
+; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v5
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[10:11]
+; GFX6-NEXT: s_cbranch_execnz .LBB16_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[10:11]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(7) %ptr, i32 256
+ %unused = atomicrmw fmin ptr addrspace(7) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret void
+}
+
+define <2 x bfloat> @buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset__waterfall(ptr addrspace(7) %ptr, <2 x bfloat> %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset__waterfall:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX12-NEXT: s_mov_b32 s1, exec_lo
+; GFX12-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: buffer_load_b32 v6, v4, s[4:7], null offen offset:1024
+; GFX12-NEXT: ; implicit-def: $vgpr4
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB17_1
+; GFX12-NEXT: ; %bb.2:
+; GFX12-NEXT: s_mov_b32 exec_lo, s1
+; GFX12-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX12-NEXT: v_and_b32_e32 v9, 0xffff0000, v5
+; GFX12-NEXT: s_mov_b32 s1, 0
+; GFX12-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Loop Header: Depth=1
+; GFX12-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX12-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX12-NEXT: s_mov_b32 s2, exec_lo
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_dual_min_num_f32 v5, v5, v9 :: v_dual_min_num_f32 v4, v4, v8
+; GFX12-NEXT: v_bfe_u32 v11, v5, 16, 1
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_bfe_u32 v10, v4, 16, 1
+; GFX12-NEXT: v_or_b32_e32 v12, 0x400000, v4
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX12-NEXT: v_or_b32_e32 v13, 0x400000, v5
+; GFX12-NEXT: v_add3_u32 v11, v11, v5, 0x7fff
+; GFX12-NEXT: v_add3_u32 v10, v10, v4, 0x7fff
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX12-NEXT: v_cndmask_b32_e32 v4, v10, v12, vcc_lo
+; GFX12-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX12-NEXT: v_cndmask_b32_e32 v5, v11, v13, vcc_lo
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_perm_b32 v5, v5, v4, 0x7060302
+; GFX12-NEXT: v_mov_b32_e32 v4, v5
+; GFX12-NEXT: v_mov_b32_e32 v5, v6
+; GFX12-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX12-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX12-NEXT: v_readfirstlane_b32 s4, v0
+; GFX12-NEXT: v_readfirstlane_b32 s5, v1
+; GFX12-NEXT: v_readfirstlane_b32 s6, v2
+; GFX12-NEXT: v_readfirstlane_b32 s7, v3
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX12-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX12-NEXT: s_and_saveexec_b32 s0, s0
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX12-NEXT: s_cbranch_execnz .LBB17_4
+; GFX12-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX12-NEXT: s_mov_b32 exec_lo, s2
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX12-NEXT: v_mov_b32_e32 v6, v4
+; GFX12-NEXT: global_inv scope:SCOPE_DEV
+; GFX12-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: s_cbranch_execnz .LBB17_3
+; GFX12-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX12-NEXT: v_mov_b32_e32 v0, v4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset__waterfall:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_add_u32_e32 v8, 0x400, v4
+; GFX940-NEXT: s_mov_b64 s[2:3], exec
+; GFX940-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: buffer_load_dword v7, v4, s[4:7], 0 offen offset:1024
+; GFX940-NEXT: ; implicit-def: $vgpr4
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB17_1
+; GFX940-NEXT: ; %bb.2:
+; GFX940-NEXT: s_mov_b64 exec, s[2:3]
+; GFX940-NEXT: s_mov_b64 s[2:3], 0
+; GFX940-NEXT: v_lshlrev_b32_e32 v9, 16, v5
+; GFX940-NEXT: s_movk_i32 s10, 0x7fff
+; GFX940-NEXT: v_and_b32_e32 v10, 0xffff0000, v5
+; GFX940-NEXT: s_mov_b32 s11, 0x7060302
+; GFX940-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Loop Header: Depth=1
+; GFX940-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_lshlrev_b32_e32 v4, 16, v7
+; GFX940-NEXT: v_min_f32_e32 v4, v4, v9
+; GFX940-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX940-NEXT: v_add3_u32 v5, v5, v4, s10
+; GFX940-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX940-NEXT: s_mov_b64 s[8:9], exec
+; GFX940-NEXT: buffer_wbl2 sc1
+; GFX940-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX940-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX940-NEXT: v_min_f32_e32 v5, v5, v10
+; GFX940-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX940-NEXT: v_add3_u32 v6, v6, v5, s10
+; GFX940-NEXT: v_or_b32_e32 v11, 0x400000, v5
+; GFX940-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX940-NEXT: s_nop 1
+; GFX940-NEXT: v_cndmask_b32_e32 v5, v6, v11, vcc
+; GFX940-NEXT: v_perm_b32 v6, v5, v4, s11
+; GFX940-NEXT: v_mov_b64_e32 v[4:5], v[6:7]
+; GFX940-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX940-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX940-NEXT: v_readfirstlane_b32 s4, v0
+; GFX940-NEXT: v_readfirstlane_b32 s5, v1
+; GFX940-NEXT: v_readfirstlane_b32 s6, v2
+; GFX940-NEXT: v_readfirstlane_b32 s7, v3
+; GFX940-NEXT: v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
+; GFX940-NEXT: s_and_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT: s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[4:7], 0 offen sc0
+; GFX940-NEXT: s_xor_b64 exec, exec, s[0:1]
+; GFX940-NEXT: s_cbranch_execnz .LBB17_4
+; GFX940-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX940-NEXT: s_mov_b64 exec, s[8:9]
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX940-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v7, v4
+; GFX940-NEXT: buffer_inv sc1
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT: s_cbranch_execnz .LBB17_3
+; GFX940-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[2:3]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset__waterfall:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX11-NEXT: s_mov_b32 s1, 0
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: buffer_load_b32 v6, v4, s[4:7], 0 offen offset:1024
+; GFX11-NEXT: ; implicit-def: $vgpr4
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB17_1
+; GFX11-NEXT: ; %bb.2:
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v5
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x1
+; GFX11-NEXT: .p2align 6
+; GFX11-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Loop Header: Depth=1
+; GFX11-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX11-NEXT: s_mov_b32 s2, exec_lo
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_dual_min_f32 v5, v5, v9 :: v_dual_min_f32 v4, v4, v8
+; GFX11-NEXT: v_bfe_u32 v11, v5, 16, 1
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_bfe_u32 v10, v4, 16, 1
+; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v4
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v5
+; GFX11-NEXT: v_add3_u32 v11, v11, v5, 0x7fff
+; GFX11-NEXT: v_add3_u32 v10, v10, v4, 0x7fff
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT: v_cndmask_b32_e32 v4, v10, v12, vcc_lo
+; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT: v_cndmask_b32_e32 v5, v11, v13, vcc_lo
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_perm_b32 v5, v5, v4, 0x7060302
+; GFX11-NEXT: v_mov_b32_e32 v4, v5
+; GFX11-NEXT: v_mov_b32_e32 v5, v6
+; GFX11-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX11-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX11-NEXT: v_readfirstlane_b32 s4, v0
+; GFX11-NEXT: v_readfirstlane_b32 s5, v1
+; GFX11-NEXT: v_readfirstlane_b32 s6, v2
+; GFX11-NEXT: v_readfirstlane_b32 s7, v3
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, vcc_lo, s0
+; GFX11-NEXT: s_and_saveexec_b32 s0, s0
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v7, s[4:7], 0 offen glc
+; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; GFX11-NEXT: s_cbranch_execnz .LBB17_4
+; GFX11-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX11-NEXT: s_mov_b32 exec_lo, s2
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX11-NEXT: v_mov_b32_e32 v6, v4
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: s_or_b32 s1, vcc_lo, s1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: s_cbranch_execnz .LBB17_3
+; GFX11-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX11-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, v4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset__waterfall:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_add_nc_u32_e32 v7, 0x400, v4
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX10-NEXT: ; implicit-def: $vgpr4
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB17_1
+; GFX10-NEXT: ; %bb.2:
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX10-NEXT: v_and_b32_e32 v9, 0xffff0000, v5
+; GFX10-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Loop Header: Depth=1
+; GFX10-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX10-NEXT: s_mov_b32 s6, exec_lo
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_min_f32_e32 v4, v4, v8
+; GFX10-NEXT: v_min_f32_e32 v5, v5, v9
+; GFX10-NEXT: v_bfe_u32 v10, v4, 16, 1
+; GFX10-NEXT: v_bfe_u32 v11, v5, 16, 1
+; GFX10-NEXT: v_or_b32_e32 v12, 0x400000, v4
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX10-NEXT: v_or_b32_e32 v13, 0x400000, v5
+; GFX10-NEXT: v_add3_u32 v10, v10, v4, 0x7fff
+; GFX10-NEXT: v_add3_u32 v11, v11, v5, 0x7fff
+; GFX10-NEXT: v_cndmask_b32_e32 v4, v10, v12, vcc_lo
+; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX10-NEXT: v_cndmask_b32_e32 v5, v11, v13, vcc_lo
+; GFX10-NEXT: v_perm_b32 v5, v5, v4, 0x7060302
+; GFX10-NEXT: v_mov_b32_e32 v4, v5
+; GFX10-NEXT: v_mov_b32_e32 v5, v6
+; GFX10-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX10-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX10-NEXT: v_readfirstlane_b32 s8, v0
+; GFX10-NEXT: v_readfirstlane_b32 s9, v1
+; GFX10-NEXT: v_readfirstlane_b32 s10, v2
+; GFX10-NEXT: v_readfirstlane_b32 s11, v3
+; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[0:1]
+; GFX10-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[2:3]
+; GFX10-NEXT: s_and_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_and_saveexec_b32 s4, s4
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_xor_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB17_4
+; GFX10-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX10-NEXT: s_mov_b32 exec_lo, s6
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6
+; GFX10-NEXT: v_mov_b32_e32 v6, v4
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5
+; GFX10-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: s_cbranch_execnz .LBB17_3
+; GFX10-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s5
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset__waterfall:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_add_u32_e32 v8, 0x400, v4
+; GFX90A-NEXT: s_mov_b64 s[6:7], exec
+; GFX90A-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_nop 0
+; GFX90A-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen offset:1024
+; GFX90A-NEXT: ; implicit-def: $vgpr4
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB17_1
+; GFX90A-NEXT: ; %bb.2:
+; GFX90A-NEXT: s_mov_b64 exec, s[6:7]
+; GFX90A-NEXT: s_mov_b64 s[6:7], 0
+; GFX90A-NEXT: v_lshlrev_b32_e32 v9, 16, v5
+; GFX90A-NEXT: s_movk_i32 s14, 0x7fff
+; GFX90A-NEXT: v_and_b32_e32 v10, 0xffff0000, v5
+; GFX90A-NEXT: s_mov_b32 s15, 0x7060302
+; GFX90A-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Loop Header: Depth=1
+; GFX90A-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v7
+; GFX90A-NEXT: v_min_f32_e32 v4, v4, v9
+; GFX90A-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX90A-NEXT: v_add3_u32 v5, v5, v4, s14
+; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v4
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX90A-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
+; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
+; GFX90A-NEXT: v_min_f32_e32 v5, v5, v10
+; GFX90A-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX90A-NEXT: v_add3_u32 v6, v6, v5, s14
+; GFX90A-NEXT: v_or_b32_e32 v11, 0x400000, v5
+; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX90A-NEXT: v_cndmask_b32_e32 v5, v6, v11, vcc
+; GFX90A-NEXT: v_perm_b32 v6, v5, v4, s15
+; GFX90A-NEXT: s_mov_b64 s[12:13], exec
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[6:7] op_sel:[0,1]
+; GFX90A-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX90A-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX90A-NEXT: v_readfirstlane_b32 s8, v0
+; GFX90A-NEXT: v_readfirstlane_b32 s9, v1
+; GFX90A-NEXT: v_readfirstlane_b32 s10, v2
+; GFX90A-NEXT: v_readfirstlane_b32 s11, v3
+; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX90A-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX90A-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX90A-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_atomic_cmpswap v[4:5], v8, s[8:11], 0 offen glc
+; GFX90A-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX90A-NEXT: s_cbranch_execnz .LBB17_4
+; GFX90A-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX90A-NEXT: s_mov_b64 exec, s[12:13]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v7
+; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v7, v4
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: s_cbranch_execnz .LBB17_3
+; GFX90A-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset__waterfall:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_add_u32_e32 v7, 0x400, v4
+; GFX908-NEXT: s_mov_b64 s[6:7], exec
+; GFX908-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_nop 0
+; GFX908-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX908-NEXT: ; implicit-def: $vgpr4
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_1
+; GFX908-NEXT: ; %bb.2:
+; GFX908-NEXT: s_mov_b64 exec, s[6:7]
+; GFX908-NEXT: s_mov_b64 s[6:7], 0
+; GFX908-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX908-NEXT: s_movk_i32 s14, 0x7fff
+; GFX908-NEXT: v_and_b32_e32 v9, 0xffff0000, v5
+; GFX908-NEXT: s_mov_b32 s15, 0x7060302
+; GFX908-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Loop Header: Depth=1
+; GFX908-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX908-NEXT: v_min_f32_e32 v4, v4, v8
+; GFX908-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX908-NEXT: v_add3_u32 v5, v5, v4, s14
+; GFX908-NEXT: v_or_b32_e32 v10, 0x400000, v4
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX908-NEXT: v_cndmask_b32_e32 v4, v5, v10, vcc
+; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX908-NEXT: v_min_f32_e32 v5, v5, v9
+; GFX908-NEXT: v_bfe_u32 v10, v5, 16, 1
+; GFX908-NEXT: v_add3_u32 v10, v10, v5, s14
+; GFX908-NEXT: v_or_b32_e32 v11, 0x400000, v5
+; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX908-NEXT: v_cndmask_b32_e32 v5, v10, v11, vcc
+; GFX908-NEXT: v_perm_b32 v5, v5, v4, s15
+; GFX908-NEXT: v_mov_b32_e32 v4, v5
+; GFX908-NEXT: s_mov_b64 s[12:13], exec
+; GFX908-NEXT: v_mov_b32_e32 v5, v6
+; GFX908-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX908-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX908-NEXT: v_readfirstlane_b32 s8, v0
+; GFX908-NEXT: v_readfirstlane_b32 s9, v1
+; GFX908-NEXT: v_readfirstlane_b32 s10, v2
+; GFX908-NEXT: v_readfirstlane_b32 s11, v3
+; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX908-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX908-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX908-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX908-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_4
+; GFX908-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX908-NEXT: s_mov_b64 exec, s[12:13]
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v6, v4
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX908-NEXT: s_cbranch_execnz .LBB17_3
+; GFX908-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset__waterfall:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x400, v4
+; GFX8-NEXT: s_mov_b64 s[6:7], exec
+; GFX8-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_nop 0
+; GFX8-NEXT: buffer_load_dword v6, v4, s[8:11], 0 offen offset:1024
+; GFX8-NEXT: ; implicit-def: $vgpr4
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_1
+; GFX8-NEXT: ; %bb.2:
+; GFX8-NEXT: s_mov_b64 exec, s[6:7]
+; GFX8-NEXT: s_mov_b64 s[6:7], 0
+; GFX8-NEXT: v_lshlrev_b32_e32 v8, 16, v5
+; GFX8-NEXT: v_and_b32_e32 v9, 0xffff0000, v5
+; GFX8-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Loop Header: Depth=1
+; GFX8-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v6
+; GFX8-NEXT: v_min_f32_e32 v4, v4, v8
+; GFX8-NEXT: v_bfe_u32 v5, v4, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v4
+; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
+; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v4
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
+; GFX8-NEXT: v_cndmask_b32_e32 v4, v5, v10, vcc
+; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
+; GFX8-NEXT: v_min_f32_e32 v5, v5, v9
+; GFX8-NEXT: v_bfe_u32 v10, v5, 16, 1
+; GFX8-NEXT: v_add_u32_e32 v10, vcc, v10, v5
+; GFX8-NEXT: v_add_u32_e32 v10, vcc, 0x7fff, v10
+; GFX8-NEXT: v_or_b32_e32 v11, 0x400000, v5
+; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
+; GFX8-NEXT: v_cndmask_b32_e32 v5, v10, v11, vcc
+; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX8-NEXT: v_alignbit_b32 v5, v5, v4, 16
+; GFX8-NEXT: v_mov_b32_e32 v4, v5
+; GFX8-NEXT: s_mov_b64 s[12:13], exec
+; GFX8-NEXT: v_mov_b32_e32 v5, v6
+; GFX8-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX8-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX8-NEXT: v_readfirstlane_b32 s8, v0
+; GFX8-NEXT: v_readfirstlane_b32 s9, v1
+; GFX8-NEXT: v_readfirstlane_b32 s10, v2
+; GFX8-NEXT: v_readfirstlane_b32 s11, v3
+; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX8-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX8-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX8-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_atomic_cmpswap v[4:5], v7, s[8:11], 0 offen glc
+; GFX8-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_4
+; GFX8-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX8-NEXT: s_mov_b64 exec, s[12:13]
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v6
+; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v6, v4
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX8-NEXT: s_cbranch_execnz .LBB17_3
+; GFX8-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset__waterfall:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_add_i32_e32 v8, vcc, 0x400, v4
+; GFX7-NEXT: s_mov_b64 s[6:7], exec
+; GFX7-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen offset:1024
+; GFX7-NEXT: ; implicit-def: $vgpr4
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_1
+; GFX7-NEXT: ; %bb.2:
+; GFX7-NEXT: s_mov_b64 exec, s[6:7]
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v7
+; GFX7-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; GFX7-NEXT: s_mov_b64 s[6:7], 0
+; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v6
+; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v5
+; GFX7-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Loop Header: Depth=1
+; GFX7-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v4
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v7
+; GFX7-NEXT: v_min_f32_e32 v4, v4, v9
+; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v6
+; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX7-NEXT: v_min_f32_e32 v7, v7, v10
+; GFX7-NEXT: v_alignbit_b32 v5, v5, v6, 16
+; GFX7-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX7-NEXT: v_mov_b32_e32 v7, v5
+; GFX7-NEXT: s_mov_b64 s[12:13], exec
+; GFX7-NEXT: v_mov_b32_e32 v6, v4
+; GFX7-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX7-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX7-NEXT: v_readfirstlane_b32 s8, v0
+; GFX7-NEXT: v_readfirstlane_b32 s9, v1
+; GFX7-NEXT: v_readfirstlane_b32 s10, v2
+; GFX7-NEXT: v_readfirstlane_b32 s11, v3
+; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX7-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX7-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX7-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_atomic_cmpswap v[6:7], v8, s[8:11], 0 offen glc
+; GFX7-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_4
+; GFX7-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX7-NEXT: s_mov_b64 exec, s[12:13]
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX7-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX7-NEXT: v_lshlrev_b32_e32 v7, 16, v6
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX7-NEXT: s_cbranch_execnz .LBB17_3
+; GFX7-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX7-NEXT: v_mov_b32_e32 v0, v7
+; GFX7-NEXT: v_mov_b32_e32 v1, v4
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_v2bf16__offset__waterfall:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_add_i32_e32 v8, vcc, 0x400, v4
+; GFX6-NEXT: s_mov_b64 s[6:7], exec
+; GFX6-NEXT: .LBB17_1: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: buffer_load_dword v7, v4, s[8:11], 0 offen offset:1024
+; GFX6-NEXT: ; implicit-def: $vgpr4
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB17_1
+; GFX6-NEXT: ; %bb.2:
+; GFX6-NEXT: s_mov_b64 exec, s[6:7]
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v6
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v5
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v7
+; GFX6-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; GFX6-NEXT: s_mov_b64 s[6:7], 0
+; GFX6-NEXT: v_and_b32_e32 v9, 0xffff0000, v6
+; GFX6-NEXT: v_and_b32_e32 v10, 0xffff0000, v5
+; GFX6-NEXT: .LBB17_3: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Loop Header: Depth=1
+; GFX6-NEXT: ; Child Loop BB17_4 Depth 2
+; GFX6-NEXT: v_mul_f32_e32 v5, 1.0, v4
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v5
+; GFX6-NEXT: v_mul_f32_e32 v6, 1.0, v7
+; GFX6-NEXT: v_min_f32_e32 v4, v4, v9
+; GFX6-NEXT: v_and_b32_e32 v7, 0xffff0000, v6
+; GFX6-NEXT: v_lshrrev_b32_e32 v5, 16, v5
+; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4
+; GFX6-NEXT: v_min_f32_e32 v7, v7, v10
+; GFX6-NEXT: v_alignbit_b32 v5, v5, v6, 16
+; GFX6-NEXT: v_alignbit_b32 v4, v4, v7, 16
+; GFX6-NEXT: v_mov_b32_e32 v7, v5
+; GFX6-NEXT: s_mov_b64 s[12:13], exec
+; GFX6-NEXT: v_mov_b32_e32 v6, v4
+; GFX6-NEXT: .LBB17_4: ; Parent Loop BB17_3 Depth=1
+; GFX6-NEXT: ; => This Inner Loop Header: Depth=2
+; GFX6-NEXT: v_readfirstlane_b32 s8, v0
+; GFX6-NEXT: v_readfirstlane_b32 s9, v1
+; GFX6-NEXT: v_readfirstlane_b32 s10, v2
+; GFX6-NEXT: v_readfirstlane_b32 s11, v3
+; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GFX6-NEXT: v_cmp_eq_u64_e64 s[4:5], s[10:11], v[2:3]
+; GFX6-NEXT: s_and_b64 s[4:5], vcc, s[4:5]
+; GFX6-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GFX6-NEXT: buffer_atomic_cmpswap v[6:7], v8, s[8:11], 0 offen glc
+; GFX6-NEXT: s_xor_b64 exec, exec, s[4:5]
+; GFX6-NEXT: s_cbranch_execnz .LBB17_4
+; GFX6-NEXT: ; %bb.5: ; in Loop: Header=BB17_3 Depth=1
+; GFX6-NEXT: s_mov_b64 exec, s[12:13]
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v6, v5
+; GFX6-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
+; GFX6-NEXT: s_or_b64 s[6:7], vcc, s[6:7]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_lshlrev_b32_e32 v7, 16, v6
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[6:7]
+; GFX6-NEXT: s_cbranch_execnz .LBB17_3
+; GFX6-NEXT: ; %bb.6: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[6:7]
+; GFX6-NEXT: v_mov_b32_e32 v0, v7
+; GFX6-NEXT: v_mov_b32_e32 v1, v4
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr <2 x bfloat>, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmin ptr addrspace(7) %gep, <2 x bfloat> %val syncscope("agent") seq_cst
+ ret <2 x bfloat> %result
+}
+
+; --------------------------------------------------------------------
+; misc
+; --------------------------------------------------------------------
+
+define float @buffer_fat_ptr_system_atomic_fmin_ret_f32__offset(ptr addrspace(7) inreg %ptr, float %val) #0 {
+; GFX12-LABEL: buffer_fat_ptr_system_atomic_fmin_ret_f32__offset:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX12-NEXT: s_wait_expcnt 0x0
+; GFX12-NEXT: s_wait_samplecnt 0x0
+; GFX12-NEXT: s_wait_bvhcnt 0x0
+; GFX12-NEXT: s_wait_kmcnt 0x0
+; GFX12-NEXT: v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s4
+; GFX12-NEXT: s_addk_co_i32 s4, 0x400
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT: v_dual_mov_b32 v3, s4 :: v_dual_max_num_f32 v2, v1, v1
+; GFX12-NEXT: buffer_load_b32 v0, v0, s[0:3], null offen offset:1024
+; GFX12-NEXT: s_mov_b32 s4, 0
+; GFX12-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: v_mov_b32_e32 v5, v0
+; GFX12-NEXT: s_wait_storecnt 0x0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-NEXT: v_max_num_f32_e32 v0, v5, v5
+; GFX12-NEXT: v_min_num_f32_e32 v4, v0, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX12-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v3, s[0:3], null offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT: s_wait_loadcnt 0x0
+; GFX12-NEXT: global_inv scope:SCOPE_SYS
+; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX12-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_cbranch_execnz .LBB18_1
+; GFX12-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX12-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX12-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX940-LABEL: buffer_fat_ptr_system_atomic_fmin_ret_f32__offset:
+; GFX940: ; %bb.0:
+; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v1, v0
+; GFX940-NEXT: v_mov_b32_e32 v0, s4
+; GFX940-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen offset:1024
+; GFX940-NEXT: s_add_i32 s6, s4, 0x400
+; GFX940-NEXT: s_mov_b64 s[4:5], 0
+; GFX940-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX940-NEXT: v_mov_b32_e32 v3, s6
+; GFX940-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX940-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: v_mov_b32_e32 v5, v0
+; GFX940-NEXT: v_max_f32_e32 v0, v5, v5
+; GFX940-NEXT: v_min_f32_e32 v4, v0, v2
+; GFX940-NEXT: v_mov_b64_e32 v[0:1], v[4:5]
+; GFX940-NEXT: buffer_wbl2 sc0 sc1
+; GFX940-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[0:3], 0 offen sc0
+; GFX940-NEXT: s_waitcnt vmcnt(0)
+; GFX940-NEXT: buffer_inv sc0 sc1
+; GFX940-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX940-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX940-NEXT: s_andn2_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_cbranch_execnz .LBB18_1
+; GFX940-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT: s_or_b64 exec, exec, s[4:5]
+; GFX940-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-LABEL: buffer_fat_ptr_system_atomic_fmin_ret_f32__offset:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT: v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s4
+; GFX11-NEXT: s_addk_i32 s4, 0x400
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: v_dual_mov_b32 v3, s4 :: v_dual_max_f32 v2, v1, v1
+; GFX11-NEXT: buffer_load_b32 v0, v0, s[0:3], 0 offen offset:1024
+; GFX11-NEXT: s_mov_b32 s4, 0
+; GFX11-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: v_mov_b32_e32 v5, v0
+; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_max_f32_e32 v0, v5, v5
+; GFX11-NEXT: v_min_f32_e32 v4, v0, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v5
+; GFX11-NEXT: buffer_atomic_cmpswap_b32 v[0:1], v3, s[0:3], 0 offen glc
+; GFX11-NEXT: s_waitcnt vmcnt(0)
+; GFX11-NEXT: buffer_gl1_inv
+; GFX11-NEXT: buffer_gl0_inv
+; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX11-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_cbranch_execnz .LBB18_1
+; GFX11-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX11-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: buffer_fat_ptr_system_atomic_fmin_ret_f32__offset:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v1, v0
+; GFX10-NEXT: v_mov_b32_e32 v0, s8
+; GFX10-NEXT: s_addk_i32 s8, 0x400
+; GFX10-NEXT: v_mov_b32_e32 v3, s8
+; GFX10-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX10-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX10-NEXT: s_mov_b32 s8, 0
+; GFX10-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: v_mov_b32_e32 v5, v0
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_max_f32_e32 v0, v5, v5
+; GFX10-NEXT: v_min_f32_e32 v4, v0, v2
+; GFX10-NEXT: v_mov_b32_e32 v0, v4
+; GFX10-NEXT: v_mov_b32_e32 v1, v5
+; GFX10-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX10-NEXT: s_waitcnt vmcnt(0)
+; GFX10-NEXT: buffer_gl1_inv
+; GFX10-NEXT: buffer_gl0_inv
+; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v5
+; GFX10-NEXT: s_or_b32 s8, vcc_lo, s8
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_cbranch_execnz .LBB18_1
+; GFX10-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s8
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX90A-LABEL: buffer_fat_ptr_system_atomic_fmin_ret_f32__offset:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v1, v0
+; GFX90A-NEXT: v_mov_b32_e32 v0, s8
+; GFX90A-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX90A-NEXT: s_add_i32 s10, s8, 0x400
+; GFX90A-NEXT: s_mov_b64 s[8:9], 0
+; GFX90A-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX90A-NEXT: v_mov_b32_e32 v3, s10
+; GFX90A-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_mov_b32_e32 v5, v0
+; GFX90A-NEXT: v_max_f32_e32 v0, v5, v5
+; GFX90A-NEXT: v_min_f32_e32 v4, v0, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[4:5] op_sel:[0,1]
+; GFX90A-NEXT: buffer_wbl2
+; GFX90A-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: buffer_invl2
+; GFX90A-NEXT: buffer_wbinvl1
+; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX90A-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX90A-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_cbranch_execnz .LBB18_1
+; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX90A-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX908-LABEL: buffer_fat_ptr_system_atomic_fmin_ret_f32__offset:
+; GFX908: ; %bb.0:
+; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v1, v0
+; GFX908-NEXT: v_mov_b32_e32 v0, s8
+; GFX908-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX908-NEXT: s_add_i32 s10, s8, 0x400
+; GFX908-NEXT: s_mov_b64 s[8:9], 0
+; GFX908-NEXT: v_max_f32_e32 v2, v1, v1
+; GFX908-NEXT: v_mov_b32_e32 v3, s10
+; GFX908-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: v_mov_b32_e32 v5, v0
+; GFX908-NEXT: v_max_f32_e32 v0, v5, v5
+; GFX908-NEXT: v_min_f32_e32 v4, v0, v2
+; GFX908-NEXT: v_mov_b32_e32 v0, v4
+; GFX908-NEXT: v_mov_b32_e32 v1, v5
+; GFX908-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX908-NEXT: s_waitcnt vmcnt(0)
+; GFX908-NEXT: buffer_wbinvl1
+; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX908-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX908-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_cbranch_execnz .LBB18_1
+; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX908-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX908-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: buffer_fat_ptr_system_atomic_fmin_ret_f32__offset:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v1, v0
+; GFX8-NEXT: v_mov_b32_e32 v0, s8
+; GFX8-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX8-NEXT: s_add_i32 s10, s8, 0x400
+; GFX8-NEXT: s_mov_b64 s[8:9], 0
+; GFX8-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX8-NEXT: v_mov_b32_e32 v3, s10
+; GFX8-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v5, v0
+; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v5
+; GFX8-NEXT: v_min_f32_e32 v4, v0, v2
+; GFX8-NEXT: v_mov_b32_e32 v0, v4
+; GFX8-NEXT: v_mov_b32_e32 v1, v5
+; GFX8-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: buffer_wbinvl1
+; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX8-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX8-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_cbranch_execnz .LBB18_1
+; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX8-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: buffer_fat_ptr_system_atomic_fmin_ret_f32__offset:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v1, v0
+; GFX7-NEXT: v_mov_b32_e32 v0, s8
+; GFX7-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX7-NEXT: s_add_i32 s10, s8, 0x400
+; GFX7-NEXT: s_mov_b64 s[8:9], 0
+; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX7-NEXT: v_mov_b32_e32 v3, s10
+; GFX7-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v5, v0
+; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v5
+; GFX7-NEXT: v_min_f32_e32 v4, v0, v2
+; GFX7-NEXT: v_mov_b32_e32 v0, v4
+; GFX7-NEXT: v_mov_b32_e32 v1, v5
+; GFX7-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: buffer_wbinvl1
+; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX7-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX7-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_cbranch_execnz .LBB18_1
+; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX7-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: buffer_fat_ptr_system_atomic_fmin_ret_f32__offset:
+; GFX6: ; %bb.0:
+; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v1, v0
+; GFX6-NEXT: v_mov_b32_e32 v0, s8
+; GFX6-NEXT: buffer_load_dword v0, v0, s[4:7], 0 offen offset:1024
+; GFX6-NEXT: s_add_i32 s10, s8, 0x400
+; GFX6-NEXT: s_mov_b64 s[8:9], 0
+; GFX6-NEXT: v_mul_f32_e32 v2, 1.0, v1
+; GFX6-NEXT: v_mov_b32_e32 v3, s10
+; GFX6-NEXT: .LBB18_1: ; %atomicrmw.start
+; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: v_mov_b32_e32 v5, v0
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: v_mul_f32_e32 v0, 1.0, v5
+; GFX6-NEXT: v_min_f32_e32 v4, v0, v2
+; GFX6-NEXT: v_mov_b32_e32 v0, v4
+; GFX6-NEXT: v_mov_b32_e32 v1, v5
+; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], v3, s[4:7], 0 offen glc
+; GFX6-NEXT: s_waitcnt vmcnt(0)
+; GFX6-NEXT: buffer_wbinvl1
+; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX6-NEXT: s_or_b64 s[8:9], vcc, s[8:9]
+; GFX6-NEXT: s_andn2_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_cbranch_execnz .LBB18_1
+; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end
+; GFX6-NEXT: s_or_b64 exec, exec, s[8:9]
+; GFX6-NEXT: s_waitcnt expcnt(0)
+; GFX6-NEXT: s_setpc_b64 s[30:31]
+ %gep = getelementptr float, ptr addrspace(7) %ptr, i32 256
+ %result = atomicrmw fmin ptr addrspace(7) %gep, float %val seq_cst
+ ret float %result
+}
+
+attributes #0 = { nounwind "amdgpu-unsafe-fp-atomics"="true" }
+
+
More information about the llvm-commits
mailing list